# ELPTNet: An Efficient LiDAR-based 3D Pedestrian Tracking Network for Autonomous Navigation Social Robots Jinzheng Guang, Zhenzhong Cao, Yinuo Song, and Jingtai Liu $^{*}$, Senior Member, IEEE Abstract— Autonomous navigation social robots need to track pedestrian movements in real-time with high precision to optimize path planning and avoid collisions. However, the main challenge of pedestrian tracking lies in the significant variations in human posture, which differ from rigid-body structures like vehicles. In this paper, we propose an Efficient LiDAR-based 3D Pedestrian Tracking Network (ELPTNet). First, our ELPTNet employs a 3D object detector to extract directional 3D pedestrian bounding boxes from LiDAR point clouds. Then, our ELPTNet employs a Constant Acceleration (CA) model and prediction confidence for target trajectory prediction. During the data association process, it integrates geometric, appearance, and motion features to enhance the robustness and real-time performance of 3D MOT when targets are temporarily occluded. Experimental results demonstrate that our ELPTNet achieves the highest ranking on the large-scale JRDB dataset for the 3D tracking task, outperforming previous state-of-the-art (SOTA) methods with improvements of 8.4% in MOTA and 6.6% in HOTA. Additionally, our ELPTNet attains an inference speed of 61 frames per second (FPS) on a single CPU. Therefore, our method enables accurate and real-time tracking of multiple pedestrians. The code is publicly available at . ### I. INTRODUCTION With the rapid advancement of artificial intelligence technology [^1], [^2], [^3], pedestrian tracking has become a critical task across various domains, such as autonomous driving [^4], [^5] and mobile robotics [^6], [^7]. In autonomous navigation social robot systems, accurate pedestrian localization and tracking are key to achieving autonomous navigation and providing comfortable services [^8]. However, traditional image-based pedestrian tracking methods [^9] are highly sensitive to lighting conditions, showing poor stability under low light and bright light environments, and lack depth perception capabilities. In contrast, LiDAR-based pedestrian tracking methods [^10], [^11], [^12] can effectively overcome these limitations. LiDAR generates high-density 3D point cloud data by emitting laser pulses and receiving reflected signals, providing precise spatial information, thus enabling high-precision pedestrian localization and tracking in dynamic environments, offering significant advantages. However, LiDAR-based pedestrian tracking methods still face numerous challenges. As illustrated in Fig. 1, one of the primary challenges is the significant variation in pedestrian postures over time. This variation is not only reflected in the differences in posture among different pedestrians at a given moment but also in the fluctuations in the posture of the same pedestrian over different time intervals. Another challenge lies in the sparsity of LiDAR point cloud data, especially for small-scale pedestrian targets. Thus, the accuracy of pedestrian tracking using LiDAR remains low. Regarding the aforementioned issues, we propose an Efficient LiDAR-based 3D Pedestrian Tracking Network (ELPTNet) for autonomous navigation social robots. Our ELPTNet is an online approach primarily composed of four key components: 3D pedestrian detection, state prediction, data association, and the state update module. Experimental results indicate that our ELPTNet surpasses the existing state-of-the-art (SOTA) approaches in both accuracy and efficiency. The three contributions of this paper are as follows: 1) We propose ELPTNet, a real-time online 3D MOT method that utilizes LiDAR 3D point cloud data for pedestrian tracking, effectively reducing interference caused by lighting variations. 2) Our ELPTNet employs a Constant Acceleration (CA) model and prediction confidence for target trajectory prediction. During the data association process, it integrates geometric, appearance, and motion features to enhance the robustness and real-time performance of 3D MOT when targets are temporarily occluded. 3) Experimental results demonstrate that our ELPTNet achieves SOTA performance on the large-scale JRDB dataset for the 3D tracking task. Compared to the second-best method, our ELPTNet improves MOTA and HOTA performance by 8.4% and 6.6%, respectively. ### II. RELATED WORK In recent years, 3D Multi-Object Tracking (MOT) has been categorized into two types based on the sensor used: Camera-based 2D MOT and LiDAR-based 3D MOT methods. ### A. Camera-based 2D Multi-Object Tracking Methods Most current 2D Multiple Object Tracking (MOT) methods follow a detection-based tracking framework, where data association is typically addressed through global optimization or online bipartite graph matching methods [^13]. Global optimization methods, such as network flow [^14], [^15] and Bayesian filtering [^16], optimize using information from all frames. In contrast, online bipartite graph matching methods rely only on information from the current frame to track targets, employing the Hungarian algorithm [^17], [^18] or greedy assignment algorithms [^19]. Despite the success of many 2D MOT approaches, their ability to represent the dynamics of --- ![Image](images/4b7d001db0b14c478b3d8462222d3bfc39a86909df727145fbe12ea9b152f3a8.jpg)
Fig. 1. The main challenge in pedestrian tracking is the significant variation in pedestrian postures. First, unlike rigid bodies (e.g., vehicles), the same individual (e.g., Ped 1) may exhibit different postures or states at different times, such as standing, bending, or sitting. Second, at the same time instant, different pedestrians may be in different postures, such as standing, walking, or bending over. Additionally, pedestrians vary in size, such as adults (e.g., Ped 2) and children (e.g., Ped 3). Here, “tressider-2019-03-16_2_000030” represents the 30th frame (000030) of the tressider-2019-03-16_2_scene.
3D scenes remains limited, leading to suboptimal 3D layout estimation in practical applications. ### B. LiDAR-based 3D Multi-Object Tracking Methods Most 3D MOT methods [^20], [^21], [^22] follow a workflow similar to 2D MOT, adhering to the “detection-association-state estimation” paradigm. The main difference lies in extending the bounding boxes from the image plane into three-dimensional space, which necessitates more precise modeling and management of objects’ 3D dimensions, orientations, and motion states. For example, SimpleTrack [^23] emphasizes a unified analysis and improvement of association, motion models, and lifecycle management. Meanwhile, Polymot [^24] adopted different motion models and similarity metrics for various object categories in data association and state estimation, reducing missed associations via a “polyhedral” strategy. Lastly, SMAT [^25] proposed a self-reinforcing framework that deeply couples dynamic object detection with static mapping in large-scale open environments for simultaneous localization, mapping, and tracking. In summary, although 3D MOT and 2D MOT share the same fundamental components, the more abundant geometric and motion information in three-dimensional scenarios presents new challenges and opportunities for data association and trajectory management. In addition, several methods that fuse camera and LiDAR data for multi-object tracking [^26] have also achieved promising performance, such as CAMO-MOT [^27], JRMOT [^8], and DeepFusionMOT [^28]. ### III. METHOD Our pedestrian tracking method is an online approach primarily composed of four components: a 3D pedestrian detection module, a state prediction module, a data association module, and a state update module. ### A. 3D Detection As illustrated in Fig. 2, to achieve better tracking performance, we employed a high-quality pre-trained DCCLA network [^29] as our 3D detection module, with the DCCLA method demonstrating SOTA performance on the 3D object detection task on the JRDB dataset. At time t, the 3D detection module outputs a set of detection results $P = \{P^1, P^2, \cdots, P^n\}$, where n denotes the number of detected pedestrians. Since the number of detected pedestrians varies at different times, the value of n will also change accordingly. At frame t, each pedestrian detection result can be represented as a tuple $P_t^i = (x, y, z, l, w, h, \theta, s)$, where $(x, y, z)$ indicates the 3D coordinates of the pedestrian's center, $(l, w, h)$ represents the dimensions of the pedestrian, $\theta$ denotes the heading angle, and s denotes the detection confidence. ### B. State Prediction Our adopts the Constant Acceleration (CA) model and Prediction Confidence to enhance the robustness and accuracy of 3D MOT. Unlike the traditional Constant Velocity (CV) model, which only considers velocity, the CA model further incorporates acceleration, allowing for more precise trajectory prediction when an object accelerates or decelerates. In contrast, the CV model tends to produce larger prediction errors when the object suddenly changes speed or experiences temporary occlusion. The CA model, however, adapts better to dynamic changes in motion by predicting movement using the following equations: $$ \begin{aligned}\boldsymbol{p}_{t}&=\boldsymbol{p}_{t-1}+\boldsymbol{v}_{t-1}\cdot\boldsymbol{\delta}+\frac{1}{2}\boldsymbol{a}_{t-1}\cdot\boldsymbol{\delta}^{2},\\\boldsymbol{v}_{t}&=\boldsymbol{v}_{t-1}+\boldsymbol{a}_{t-1}\cdot\boldsymbol{\delta},\end{aligned} $$ where $p_t$, $v_t$, $a_t$, and $\delta$ represent the position, velocity, acceleration of the object at time t, and the time interval between consecutive frames, respectively. Furthermore, to enhance the reliability of predictions, we propose Prediction Confidence, which assesses the uncertainty of the predicted results. The confidence level decays over time according to the following equation: $\gamma_t = \gamma_{t-1} - \eta\gamma_{t-1}$, where $\gamma_t$ represents the prediction confidence of the current frame and $\eta$ is the decay factor, which controls the rate at which confidence decreases. --- ![Image](images/68a293cdf49b46036fbb9782f6f250ef8588fe1dccf24dbfdf9d8cd4e9521fb7.jpg)
Fig. 2. Our ELPTNet pipeline. Our ELPTNet is composed of four key components: the 3D detection module, which extracts 3D bounding boxes from the LiDAR point cloud; the state prediction module, utilizing a a Constant Acceleration (CA) model and prediction confidence to estimate object states in the current frame; the data association module, responsible for linking detections with predicted trajectories; and the state update module, which refines trajectories while handling newly emerging and vanishing objects.
### C. Data Association In 3D MOT tasks, objects may become temporarily undetectable due to occlusion or long distances, leading to tracking failures. Traditional constant velocity (CV) models struggle with abrupt motion changes, while deep learning-based models (e.g., LSTMs), despite their predictive accuracy, are computationally expensive and impractical for real-time applications. To address these challenges, this paper proposes a prediction confidence-guided data association method, which introduces a prediction confidence metric $\gamma$ to quantify the reliability of predicted states and dynamically adjust the search range in data association. Specifically, our method utilizes a constant acceleration (CA) motion model for state prediction: $$ \hat{Z}_{t}^{j}=A Z_{t-1}^{j};\hat{S}_{t}^{j}=A S_{t-1}^{j}A^{T}+Q, $$ where A is the state transition matrix, $\tilde{S}_{t}$ represents the predicted error covariance, and Q is the process noise covariance. The position and velocity predictions are computed as follows: $$ \begin{align*}\hat{p}^{j}_{t}&=p^{j}_{t-1}+\delta v^{j}_{t-1}+\frac{\delta^{2}a^{j}_{t-1}}{2},\\\hat{v}^{j}_{t}&=v^{j}_{t-1}+\delta a^{j}_{t-1}.\end{align*} $$ To ensure robust data association, the prediction confidence $\gamma$ decreases over time in the absence of detection and is updated using: $$ \gamma_{t}^{j}=\gamma_{t-1}^{j}-\eta\gamma_{t-1}^{j};\gamma_{t}^{j}=\gamma_{t}^{j}+\eta c_{t}^{j}, $$ where $c_{t}^{j}$ represents the detection confidence, and $\eta$ is a parameter that controls the rate of confidence decay. The data association process is performed by computing an aggregated pairwise cost (C), which integrates geometric (G), appearance (A), and motion (M) features: $$ C^{i},j_{t}=G^{i},j_{t}+A^{i},j_{t}+M^{i},j_{t}. $$ The geometric cost (G) evaluates differences in bounding box dimensions, centroids, and orientations, while the appearance cost (A) is derived from deep features extracted by the detector and computed using the Euclidean distance. Additionally, the motion cost (M) measures the angular and Euclidean distances between velocity vectors. Ultimately, data association is performed using a cost matrix $\psi_{t}$ weighted by the prediction confidence. ### D. State Update In the target state update process, we employs Kalman Filtering to optimize the state of matched targets. First, the Kalman gain is computed to balance prediction and detection errors, updating the target state and its error covariance matrix to enhance tracking stability. Meanwhile, a Prediction Confidence mechanism is introduced to adaptively adjust the search range for data association based on detection confidence, ensuring reasonable tracking of temporarily lost targets. For unmatched detection targets, a zero-initialization strategy is adopted, adding them to the tracking state set with initial velocity and acceleration set to zero, allowing new targets to be swiftly incorporated into the tracking sequence. Ultimately, the updated target state consists of both matched targets and newly initialized targets, serving as the foundation for tracking in the next frame. By smoothing target trajectories, optimizing data association, and enhancing ---
TABLE I
COMPARISON OF TRACKING PERFORMANCE BETWEEN OUR ELPTNET AND OTHER STATE-OF-THE-ART (SOTA) METHODS ON THE JRDB TEST SET. OUR ELPTNET ACHIEVES A NEW SOTA PERFORMANCE, IMPROVING HOTA BY 6.6% COMPARED TO THE PREVIOUS SOTA METHOD [^25]. $\triangle$ REPRESENTS THE ABSOLUTE VALUE OF THE DIFFERENCE BETWEEN OUR ELPTNET AND SMAT.
MethodHOTA $\uparrow$ (%)DeTA $\uparrow$ (%)DetRe $\uparrow$ (%)DetPr $\uparrow$ (%)AssA $\uparrow$ (%)AssRe $\uparrow$ (%)AssPr $\uparrow$ (%)IDFI $\uparrow$ (%)IDR $\uparrow$ (%)IDP $\uparrow$ (%)OSPA $\downarrow$OSPACARD $\downarrow$OSPACLOC $\downarrow$
AB3DMOT [^12]11.81212.06912.55755.23412.05312.79168.36910.6706.54828.8030.9710.7340.237
JRMOT [^8]13.05312.60013.19853.73513.95614.94063.94012.3197.67231.2350.9560.6270.329
PC-DAN [^30]15.56923.68725.86370.01310.37610.77470.62913.8899.51025.7430.9760.8980.079
SS3D-MOT [^31]15.80023.22425.19670.96610.88911.18484.51914.4809.81127.6320.9820.9110.071
SMAT [^25]18.93017.05518.41755.31021.13222.34271.84221.34214.22442.7190.9750.6930.282
ELPTNet (Ours)25.53825.99128.58364.58925.42930.56053.02328.68620.69046.7530.7830.2300.553
$\triangle$6.6088.93610.1669.2794.2978.21818.8197.3446.4664.0340.1920.4630.271
TABLE II
COMPARISON OF TRACKING PERFORMANCE BETWEEN OUR ELPTNET AND OTHER SOTA METHODS ON THE JRDB TEST SET. OUR ELPTNET ACHIEVES THE TOP RANKING, IMPROVING MOTA BY 8.4% COMPARED TO THE SECOND-BEST METHOD [^25].
MethodMOTA $\uparrow$ (%)MODA $\uparrow$ (%)CLR_Re $\uparrow$ (%)CLR_Pr $\uparrow$ (%)CLR_TP $\uparrow$CLR_FN $\downarrow$CLR_FP $\downarrow$MTR $\uparrow$ (%)PTR $\downarrow$IDSW $\downarrow$MT $\uparrow$PT $\downarrow$ML $\downarrow$
AB3DMOT [^12]19.34219.96721.35093.914211186777968136864.82930.2646179865391156
JRMOT [^8]20.15220.57822.57091.890223253765901197056.06431.05042161085531120
PC-DAN [^30]22.56525.19631.06884.104307308681846580844.71745.9292602284818879
SS3D-MOT [^31]22.96324.98230.24385.182299153690001520414.21145.5361997375811895
SMAT [^25]24.21024.68928.99387.073286790702364425774.88543.627473987777917
ELPTNet (Ours)32.65333.49438.87487.8443845246046305321216.84437.2268320300663818
$\triangle$8.4438.8059.8810.77197734977341063511.9596.401358121311499
tracking robustness, this method significantly improves the accuracy and efficiency of 3D multi-object tracking. ### IV. EXPERIMENT The JRDB dataset [^32], [^8] and its evaluation metrics are first introduced. Next, the performance of ELPTNet is compared with other SOTA methods. Then, an ablation study is conducted to evaluate the performance of the method under various conditions. Finally, we present a visual analysis and discuss the robustness and limitations of the approach. ### A. Dataset and Evaluation Metric 1) Dataset: The JRDB [^32], [^8] dataset was collected by the mobile robotic arm JackRabbit while navigating and interacting in complex human environments. JackRabbit is equipped with five cameras and two 16-channel LiDAR sensors, generating 3D point cloud data that mimics a 32-channel LiDAR. This dataset provides 64 minutes of sensor data and includes over 4.2 million pedestrian 3D bounding boxes. In total, it covers 54 indoor and outdoor scenes on a university campus, of which 27 unlabeled scenes are used for the test set and the remaining 27 are used for training and validation. Following the official partitioning scheme, we select 7 of these scenes as our validation set. 2) Metric: We used the following evaluation metrics: HOTA (Higher Order Tracking Accuracy), DetA (Detection Accuracy), DetRe (Detection Recall), DetPr (Detection Precision), AssA (Association Accuracy), AssRe (Association Recall), AssPr (Association Precision), IDF1, IDR (ID Recall), IDP (ID Precision), OSPA (Optimal Sub-Pattern Assignment), OSPA_CARD (Cardinality), OSPA_LOC (Localization), MOTA (Multi-Object Tracking Accuracy), MODA (Multi-Object Detection Accuracy), CLR_Re (Clear Recall), CLR_Pr (Clear Precision), CLR_TP (Clear True Positives), CLR_FN (Clear False Negatives), CLR_FP (Clear False Positives), MTR (Mostly Tracked), PTR (Partially Tracked), MT (Mostly Tracked Targets), PT (Partially Tracked Targets), ML (Missed Targets), IDSW (ID Switches). ### B. Comparison with SOTA methods on the JRDB test set As illustrated in Table I, our ELPTNet achieved the top ranking in the 3D pedestrian tracking task on the JRDB test set $^{1}$, outperforming other methods across multiple tracking metrics. In particular, our ELPTNet attains the highest HOTA score (25.538%), outperforming the previous best method SMAT by 6.608%. It also exhibits substantial gains in both detection and association metrics: DetA, DetRe, and DetPr are improved by up to 8.936%, 10.166%, and 9.279%, respectively, whereas AssA, AssRe, and AssPr see improvements of up to 4.297%, 8.218%, and 18.819%. Furthermore, our ELPTNet yields the best IDF1, IDR, and IDP values, underscoring its effectiveness in consistently identifying tracked objects over time. In terms of overall tracking quality measured by OSPA, our ELPTNet reduces the error to 0.783, offering a notable improvement of 0.192 compared to SMAT. Similar advantages are observed in OSPA $_{CARD}$ and OSPA $_{LOC}$, demonstrating that our ELPTNet not only $^{1}$JRDB 3D Tracking: ---
TABLE III
THE TRACKING PERFORMANCE OF OUR ELPTNET AND SMAT [^25] ON THE JRDB TEST SET IS COMPARED ACROSS DIFFERENT SCENARIOS. OUR ELPTNET OUTPERFORMS THE SMAT METHOD IN MOST SCENARIOS.
SequenceELPTNet (Ours)SMAT [^25]
HOTA $\uparrow$ (%)DetA $\uparrow$ (%)AssA $\uparrow$ (%)IDF1 $\uparrow$ (%)OSPA $\downarrow$ (%)MOTA $\uparrow$ (%)MODA $\uparrow$ (%)HOTA $\uparrow$ (%)DetA $\uparrow$ (%)AssA $\uparrow$ (%)IDF1 $\uparrow$ (%)OSPA $\downarrow$ (%)MOTA $\uparrow$ (%)MODA $\uparrow$ (%)
cubberly-auditorium-2019-04-22_140.0641.4539.7047.440.6958.5959.1323.1820.8026.0625.410.9835.5535.98
discovery-walk-2019-02-28_036.1437.9534.8043.230.8449.7650.2821.1215.8330.8325.700.9626.2726.59
discovery-walk-2019-02-28_137.2836.6738.1548.560.8150.0750.3620.4918.6222.6423.560.9932.0232.36
food-trucks-2019-02-12_027.3838.8319.5529.990.9051.8252.6217.7420.5315.4521.520.9833.3033.68
gates-ai-lab-2019-04-17_025.5534.4419.0626.620.9337.0837.7823.9623.2024.7529.760.9921.6421.97
gates-basement-elevators-2019-01-17_038.5846.7831.9947.010.6858.8559.9326.3324.0730.4427.030.9640.0440.49
gates-foyer-2019-01-17_044.3649.2939.9855.900.7363.1763.3725.6231.1021.1735.370.9944.8845.27
gates-to-clark-2019-02-28_051.0246.3956.1864.550.5957.5757.6931.3627.1836.3648.970.9642.7143.06
hewlett-class-2019-01-23_034.3144.8326.6339.340.7655.4657.3927.3728.5826.5931.690.9749.0149.56
hewlett-class-2019-01-23_156.4057.1855.6973.990.4874.0674.2535.2133.9736.7638.440.9764.7365.10
huang-2-2019-01-25_143.8339.0649.6254.090.7354.1254.3723.3518.1929.9824.780.9822.8723.24
huang-intersection-2019-01-22_030.7127.2334.7240.710.8336.3936.7318.2112.9727.0618.770.9823.1523.43
indoor-coupa-cafe-2019-02-06_018.1222.8614.4519.910.9225.6626.1911.7914.429.6415.370.9915.8016.28
lomita-serra-intersection-2019-01-30_038.9533.4945.9751.440.7647.8247.8628.5818.4446.4333.480.9725.5325.86
meyer-green-2019-03-16_129.8119.2446.5631.160.8722.6422.8517.7912.1326.1912.670.9917.0617.52
nvidia-aud-2019-01-25_025.1625.2625.5830.980.8130.9731.5219.5515.5624.5720.690.9919.4619.93
nvidia-aud-2019-04-18_143.4939.2748.4850.980.7149.8650.1123.3516.8532.4026.690.9326.9127.09
nvidia-aud-2019-04-18_238.7735.2642.9048.890.7645.5945.9828.2819.9142.7830.220.9727.2627.71
outdoor-coupa-cafe-2019-02-06_017.3917.3517.7119.150.9320.2820.6113.2212.2914.3014.370.9916.5416.98
quarry-road-2019-02-28_039.5330.9950.6350.730.7637.2737.4121.8814.1333.9723.420.9820.0020.28
serra-street-2019-01-30_029.9120.4143.9134.390.8227.8528.0317.0010.0329.1415.920.9916.7117.12
stlc-111-2019-04-19_154.2557.0451.7567.320.5770.5770.9832.7432.1433.5546.200.9554.1354.57
stlc-111-2019-04-19_245.0652.1439.0557.660.6864.3464.8130.7130.6131.0145.220.9452.3952.86
tressider-2019-03-16_225.2723.4927.7931.810.8630.2230.4018.7912.8228.4021.290.9520.1720.38
tressider-2019-04-26_016.5317.2616.0920.240.9120.0021.2214.2014.8913.5519.630.9919.9520.60
tressider-2019-04-26_117.8322.5114.7519.710.9129.5830.8918.8218.3919.2821.050.9925.0025.48
tressider-2019-04-26_317.7721.3715.1519.020.8926.3627.6716.9717.4516.6419.380.9924.2424.92
COMBINED25.5425.9925.4328.690.7832.6533.4918.9317.0621.1321.340.9824.2124.69
tracks objects more accurately but also achieves more precise localization and robust object cardinality estimates. As illustrated in Table II, our ELPTNet significantly outperforms existing methods, achieving the highest Multi-Object Tracking Accuracy (MOTA) of 32.653%, which is 8.4% higher than the second-best method, SMAT [^25]. In addition, our ELPTNet attains the highest Multi-Object Detection Accuracy (MODA) of 33.494%, an 8.8% improvement over SMAT. Regarding tracking accuracy, our ELPTNet achieves the highest Correct Localization Recall (CLR_Re) of 38.874%, surpassing SMAT by 9.88%, while maintaining a competitive Correct Localization Precision (CLR_Pr) of 87.844%. Our ELPTNet also reports the highest number of CLR_TP (384,524) and the lowest number of CLR_FN (604,630) demonstrating its remarkable ability to accurately detect and track objects while minimizing missed detections. In terms of trajectory consistency metrics, our ELPTNet significantly increases the Mostly Tracked (MTR) percentage to 16.844%, outperforming all other methods. It also achieves a lower Partially Tracked (PTR) percentage of 37.226% and the fewest Mostly Lost (ML) trajectories at 818, indicating strong tracking stability. However, our ELPTNet has a slightly higher ID Switch (IDSW) count of 8,320, suggesting that identity consistency can still be improved. Overall, our ELPTNet substantially advances multi-object tracking performance. As illustrated in Table III, our ELPTNet consistently outperforms SMAT across various JRDB test sequences in key tracking metrics, including HOTA, DetA, AssA, IDF1, MOTA, and MODA, while maintaining a favorable OSPA value. Notably, the sequences “hewlett-class-2019-01-23_1” and “stlc-111-2019-04-19_1” display substantial improvements in both detection accuracy (DetA) and association accuracy (AssA), underscoring the robustness of our ELPTNet in diverse and challenging scenarios. Overall, these results highlight the effectiveness of our ELPTNet, achieving excellent tracking performance across a wide range of environments. ### C. Ablation Study We validated the impact of various components of our method on the JRDB validation set. 1) Effect of the number of new objects: As illustrated in Table IV, increasing the number of new objects generally leads to improvements in both HOTA and MOTA. Introducing 15 new objects yields the highest HOTA 21.894% and MOTA 23.690%, as well as the lowest IDSW 4341, indicating superior overall tracking accuracy and stability. Although MODA 25.525% and IDF1 23.107% achieve their respective peaks when introducing 13 and 11 new objects, respectively, the overall performance is best optimized with 15 new objects. Therefore, selecting 15 new objects is ---
TABLE IV
THE EFFECT OF THE NUMBER OF NEW OBJECTS. WHEN THE NUMBER OF NEW OBJECTS IS 15, HOTA, MOTA, AND IDSW ACHIEVE OPTIMAL PERFORMANCE.
New objectHOTA $\uparrow$ (%)MOTA $\uparrow$ (%)MODA $\uparrow$ (%)IDF1 $\uparrow$ (%)IDSW $\downarrow$
321.63523.64625.52222.8064442
521.68423.67625.52022.7444365
721.84923.64325.52322.9824451
921.61523.64525.51422.6724426
1121.75523.65625.51623.1074405
1321.86723.67525.52522.8344380
1521.89423.69025.52323.0314341
TABLE V
THE EFFECT OF THE MAXIMUM PREDICTED NEW OBJECTS COUNT. WHEN SET TO 25, IT ACHIEVES A BALANCED PERFORMANCE ACROSS MULTIPLE EVALUATION METRICS.
Max predHOTA $\uparrow$ (%)MOTA $\uparrow$ (%)MODA $\uparrow$ (%)IDF1 $\uparrow$ (%)IDSW $\downarrow$
521.70823.62325.52922.6544514
1021.76023.68625.51923.1074340
1521.64423.64025.52622.8624466
2522.03023.65825.51823.4324404
5021.95723.64925.52523.2304442
7521.94623.64525.51523.0274429
10021.97523.67325.52123.0404376
recommended to achieve better performance across multiple evaluation metrics. 2) Effect of maximum predicted new objects count: As illustrated in Table V, it can be observed that different settings for the maximum number of newly predicted objects (Max pred) produce varied effects on each metric. When Max pred is set to 25, HOTA reaches 22.030% and IDF1 reaches 23.432%, both at their highest levels. MOTA peaks at 23.686% with Max pred = 10. MODA achieves 25.529% when Max pred is set to 5, while IDSW is minimized at 4340 with Max pred = 10. These findings suggest that selecting the maximum number of newly predicted objects should balance HOTA, MOTA, MODA, IDF1, and IDSW. If HOTA and IDF1 are priorities, setting Max pred to 25 is advisable. If the focus is on MOTA or MODA, reducing Max pred to 10 or 5 may be more appropriate. It is also crucial to track changes in IDSW to minimize identity switches during the tracking process. 3) Effect of post score: As illustrated in Fig. 3, adjusting the post score threshold from 0.1 to 0.9 significantly influences tracking performance. At a threshold of 0.1, HOTA is 17.880% and IDF1 is 16.696%, indicating a high number of false positives. Increasing the threshold to 0.5 improves HOTA to 20.949% and IDF1 to 21.310%. At 0.7, HOTA rises to 21.635% and IDF1 to 22.806%, representing the best overall performance. Setting the threshold as high as 0.9, however, decreases HOTA to 14.080% and IDF1 to 15.692%, implying that many valid detections are lost. Therefore, when the post score is 0.7, both HOTA and IDF1 reach their optimal values. ![Image](images/b2263bdbdd69751c613247ffe33c5d5d6a347544e0cbd168c4f91cdf1caa82eb.jpg)
Fig. 3. The effect of the post score. When the post score is 0.7, both HOTA and IDF1 reach their highest values.
TABLE VI
THE EFFECT OF TRACKING DELAY TIME. WHEN THE LATENCY IS 0, IT INDICATES REAL-TIME ONLINE TRACKING. WHEN THE LATENCY IS 1, MEANING ONLINE TRACKING WITH A 1-SECOND DELAY, THE OVERALL TRACKING PERFORMANCE IS BETTER. THE FPS WAS TESTED ON THE INTEL CORE I9 10900K CPU.
LatencyHOTA\n $\uparrow$ (%)MOTA\n $\uparrow$ (%)MODA\n $\uparrow$ (%)IDF1\n $\uparrow$ (%)IDSW\n $\downarrow$FPS\n $\uparrow$
021.63523.64625.52222.806444261.132
122.11427.32428.18723.926204562.378
222.06726.66327.50024.035198263.681
321.96125.70226.55824.008202864.842
4) Effect of tracking delay time: As illustrated in Table VI, when the latency is 0 (real-time online tracking), the HOTA, MOTA, MODA, and IDF1 scores reach 21.635%, 23.646%, 25.522%, and 22.806%, respectively. More importantly, the frame rate achieves 61.132 FPS. As the latency increases, overall performance gradually improves. Notably, under a 1-second latency, MOTA and MODA exhibit the most significant improvements, while the IDSW count markedly decreases, and the FPS slightly increases to 62.378, resulting in the best overall tracking performance. Further increases in latency lead to continued improvements in certain metrics (e.g., IDF1 and FPS), but also cause slight declines in accuracy metrics such as MOTA and MODA, indicating that additional latency does not confer an overall benefit. In summary, our approach achieves the most balanced trade-off between computational demands and tracking accuracy, making it particularly suitable for real-time applications, including autonomous social navigation robots. 5) Robustness analysis: As illustrated in Table VII, we assessed the robustness of our multi-object tracking method in various scenarios. Despite significant differences among ---
TABLE VII
TRACKING PERFORMANCE UNDER DIFFERENT SCENARIOS. NOTE THAT OUR APPROACH DEMONSTRATES STABLE TRACKING PERFORMANCE ACROSS DIVERSE ENVIRONMENTS.
SequenceHOTA $\uparrow$ (%)DetA $\uparrow$ (%)AssA $\uparrow$ (%)MOTA $\uparrow$ (%)CLR_Re $\uparrow$ (%)CLR_Pr $\uparrow$ (%)MODA $\uparrow$ (%)PTR $\downarrow$MTR $\uparrow$ (%)IDSW $\downarrow$IDFI $\uparrow$ (%)OSPA $\downarrow$
clark-center-2019-02-28_120.29924.00817.43317.74140.46966.59620.17067.4806.504145322.0010.979
gates-ai-lab-2019-02-08_027.76727.30629.2512.31859.24351.5183.49156.00028.00012326.8150.968
huang-2-2019-01-25_030.35931.90629.02030.73353.33371.25331.81663.41519.51210737.3520.952
meyer-green-2019-03-16_034.22429.52740.01325.50852.99966.77526.62973.33313.3336540.6070.982
nvidia-aud-2019-04-18_030.33533.51027.52538.76156.12277.39139.72662.50022.50018738.2000.959
tressider-2019-03-16_125.72418.31337.02018.89327.87177.15519.61936.8427.89511624.6640.926
tressider-2019-04-26_216.75527.04910.70426.06546.20271.88828.13562.80210.145239117.5110.966
COMBINED21.63526.53518.05623.64645.36969.56725.52261.96311.861444222.8060.962
![Image](images/51aa97cdd37517cd70bf14b484e392bacc3782b8b3909ad29f340b875051b944.jpg)
Fig. 4. Visualization of meyer-green-2019-03-16_0 scene. Our ELPTNet can effectively track pedestrians. The green and red 3D bounding boxes represent the ground truth and our predicted results, respectively.
![Image](images/158487b2903ce48311fd55f6a07ed030f25834141cd27d12a7e23cce44c83453.jpg)
Fig. 5. Visualization of the tressider-2019-04-26_2 scenario. This indoor scene features high crowd density along with frequent mutual occlusions and interactions, which result in reduced tracking performance.
the scenes in terms of environmental layout, crowd density, and viewing angles, our ELPTNet demonstrated satisfactory performance in metrics such as HOTA, DetA, AssA, MOTA, MODA, and IDF1. Notably, as illustrated in Fig. 4, in the meyer-green-2019-03-16.0 scenario, due to dispersed individuals, fewer occlusions, and more stable lighting conditions, the algorithm achieved the best overall performance, with a HOTA of 34.224%, an IDF1 of 40.607%, and only 65 IDSW. Other scenarios also showed stable tracking performance. These results fully demonstrate the method's adaptability and robustness in various complex environments. ### D. Limitations As illustrated in Fig. 5, indoor scenes (e.g., tressider-2019-04-26.2) typically have high crowd density and frequent occlusions and interactions, which significantly increase the difficulty of tracking. Consequently, this scene exhibits the highest number of ID switches (2391) in Table VII, while the overall HOTA score is only 16.755%. These results indicate that although the proposed method demonstrates a certain level of robustness across various indoor and outdoor scenarios, it still encounters a large number of identity switches and tracking discontinuities in the presence of extensive occlusions and dense crowds. Therefore, further improvements and optimizations are necessary to address these challenges in complex indoor environments. ### V. CONCLUSIONS This paper presents a simple yet Efficient LiDAR-based 3D Pedestrian Tracking Network (ELPTNet) for autonomous navigation social robots. Our ELPTNet first utilizes a 3D object detector to obtain pedestrian 3D bounding boxes from LiDAR point clouds. In addition, our ELPTNet employs a Constant Acceleration (CA) model and prediction confidence for target trajectory prediction. During the data association process, it integrates geometric, appearance, and motion features to enhance the robustness and real-time performance of 3D MOT when targets are temporarily occluded. Experimental results on the large-scale JRDB pedestrian dataset --- demonstrate that our method ranks first in the JRDB 3D pedestrian tracking task, achieving an 8.4% improvement in MOTA and a 6.6% increase in HOTA compared to previous state-of-the-art (SOTA) methods. Additionally, our ELPTNet achieves an inference speed of 61 Frames Per Second (FPS) on a single CPU, enabling accurate and real-time multi-pedestrian tracking. Since LiDAR point cloud data is sparse and less accurate for tracking distant or small objects, whereas camera images excel at capturing color and fine details, future work will focus on integrating LiDAR and camera sensor data to improve tracking accuracy and robustness in complex environments for intelligent robots. ## REFERENCES [^1]: S. M. Marvasti-Zadeh, L. Cheng, H. Ghanei-Yakhdan, and S. Kasaei, “Deep learning for visual tracking: A comprehensive survey,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 5, pp. 3943–3968, 2021. [^2]: Z. Cao, Q. Zhang, J. Guang, S. Wu, Z. Hu, and J. Liu, “Semantictopoloop: Semantic loop closure with 3d topological graph based on quadric-level object map,” IEEE Robotics and Automation Letters, vol. 9, no. 5, pp. 4257–4264, 2024. [^3]: J. Guang and J. Liang, “Cmsea: Compound model scaling with efficient attention for fine-grained image classification,” IEEE Access, vol. 10, pp. 18222–18232, 2022. [^4]: M. Lyssenko, C. Gladisch, C. Heinzemann, M. Woehrle, and R. Triebel, “Towards safety-aware pedestrian detection in autonomous systems,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 293–300. [^5]: Y. Zeng, C. Ma, M. Zhu, Z. Fan, and X. Yang, “Cross-modal 3d object detection and tracking for auto-driving,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 3850–3857. [^6]: J. Guang, Z. Hu, S. Wu, Q. Zhang, and J. Liu, "Rpea: A residual path network with efficient attention for 3d pedestrian detection from lidar point clouds," Expert Systems with Applications, vol. 249, p. 123497, 2024. [^7]: D. Jia, A. Hermans, and B. Leibe, “2d vs. 3d lidar-based person detection on mobile robots,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 3604–3611. [^8]: A. Shenoi, M. Patel, J. Gwak, P. Goebel, A. Sadeghian, H. Rezatofighi, R. Martin-Martin, and S. Savarese, "Jrmot: A real-time 3d multi-object tracker and a new large-scale dataset," in 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 10335–10342. [^9]: Z. Wang, Z. Li, J. Leng, M. Li, and L. Bai, “Multiple pedestrian tracking with graph attention map on urban road scene,” IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 8, pp. 8567–8579, 2022. [^10]: Y. Yang, B. Fan, Y. Jiang, W. Zhou, D. Chen, and H. Xu, “Enhancing 3d single object tracking with efficient point cloud segmentation,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 11631–11638. [^11]: J. Shan, S. Zhou, Z. Fang, and Y. Cui, “Ptt: Point-track-transformer module for 3d single object tracking in point clouds,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 1310–1316. [^12]: X. Weng, J. Wang, D. Held, and K. Kitani, "3d multi-object tracking: A baseline and new evaluation metrics," in 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 10359–10366. [^13]: J. Wu, J. Cao, L. Song, Y. Wang, M. Yang, and J. Yuan, "Track to detect and segment: An online multi-object tracker," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 12352–12361. [^14]: P. Lenz, A. Ge: nowme: Efficient online min-cost and computation," in Computer [^15]: V. Chari, S. Lacoste-Julien, I. Laptev, and J. Sivic, “On pairwise costs for network flow multi-object tracking,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 5537–5545. [^16]: B. Lee, E. Erdenee, S. Jin, M. Y. Nam, Y. G. Jung, and P. K. Rhee, “Multi-class multi-object tracking using changing point detection,” in Computer Vision–ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14. Springer, 2016, pp. 68–83. [^17]: S. Sharma, J. A. Ansari, J. K. Murthy, and K. M. Krishna, “Beyond pixels: Leveraging geometry and shape cues for online multi-object tracking,” in 2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2018, pp. 3508–3515. [^18]: H. Karunasekera, H. Wang, and H. Zhang, “Multiple object tracking with attention to appearance, structure, motion and size,” IEEE Access, vol. 7, pp. 104423–104434, 2019. [^19]: K. Fang, Y. Xiang, X. Li, and S. Savarese, “Recurrent autoregressive networks for online multi-object tracking,” in 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). IEEE, 2018, pp. 466–475. [^20]: J. Jin, J. Zhang, K. Zhang, Y. Wang, Y. Ma, and D. Pan, "3d multi-object tracking with boosting data association and improved trajectory management mechanism," Signal Processing, vol. 218, p. 109367, 2024. [^21]: H. Wu, W. Han, C. Wen, X. Li, and C. Wang, “3d multi-object tracking in point clouds based on prediction confidence-guided data association,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 6, pp. 5668–5677, 2021. [^22]: Y. Ma, J. Zhang, G. Qin, J. Jin, K. Zhang, D. Pan, and M. Chen, "3d multi-object tracking based on dual-tracker and ds evidence theory," IEEE Transactions on Intelligent Vehicles, vol. 8, no. 3, pp. 2426–2436, 2022. [^23]: Z. Pang, Z. Li, and N. Wang, “Simpletrack: Understanding and rethinking 3d multi-object tracking,” in European Conference on Computer Vision. Springer, 2022, pp. 680–696. [^24]: X. Li, T. Xie, D. Liu, J. Gao, K. Dai, Z. Jiang, L. Zhao, and K. Wang, "Poly-mot: A polyhedral framework for 3d multi-object tracking," in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 9391–9398. [^25]: T. Fan, B. Shen, Y. Zhang, C. Zhang, L. Yang, H. Chen, W. Zhang, and J. Pan, “S2mat: Simultaneous and self-reinforced mapping and tracking in dynamic urban scenarios oring framework for simultaneous mapping and tracking in unbounded urban environments,” arXiv preprint arXiv:2304.14356, 2023. [^26]: S. Xin, Z. Zhang, L. Liu, X. Hou, D. Zhu, M. Wang, and Y. Liu, "A robotic-centric paradigm for 3d human tracking under complex environments using multi-modal adaptation," in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 4934–4940. [^27]: L. Wang, X. Zhang, W. Qin, X. Li, J. Gao, L. Yang, Z. Li, J. Li, L. Zhu, H. Wang et al., “Camo-mot: Combined appearance-motion optimization for 3d multi-object tracking with camera-lidar fusion,” IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 11, pp. 11981–11996, 2023. [^28]: X. Wang, C. Fu, Z. Li, Y. Lai, and J. He, “Deepfusionmot: A 3d multi-object tracking framework based on camera-lidar fusion with deep association,” IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 8260–8267, 2022. [^29]: J. Guang, S. Wu, Z. Hu, Q. Zhang, P. Wu, and J. Liu, “Dccla: Dense cross connections with linear attention for lidar-based 3d pedestrian detection,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 5, pp. 4535–4548, 2025. [^30]: A. Kumar, J. Kini, M. Shah, and A. Mian, "Pc-dan: Point cloud based deep affinity network for 3d multi-object tracking," arXiv preprint arXiv:2106.07552, 2021. [^31]: C. Liu, C. Gao, F. Liu, J. Liu, D. Meng, and X. Gao, “Ss3d: Sparsely-supervised 3d object detection from point cloud,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8428–8437. [^32]: R. Martin-Martin, M. Patel, H. Rezatofighi, A. Shenoi, J. Gwak, E. Frankel, A. Sadeghian, and S. Savarese, "Jrdb: A dataset and benchmark of egocentric robot visual perception of humans in built environments," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 6, pp. 6748–6765, 2023.