Spb3DTracker: A Robust LiDAR-Based Person Tracker for Noisy Environment

TL;DR

This work analyzes LiDAR-based PDT within the TBD paradigm and identifies key bottlenecks in detection post-processing, data association, motion modeling, and life-cycle management. It introduces SpbTracker, a robust TBD-based LiDAR tracker that employs a trajectory state $\mathbf{s}=[x,y,z,\theta,v_x,v_y,a_x,a_y,w,l,h]$ and a Dynamic Unscented Kalman Filter with adaptive covariance, along with a data association strategy that fuses Modified Complete-IoU (MCIoU) and BEV-based feature similarity. The system also uses a Dynamic UKF with covariance updates $S_k$ and $R_k$ to handle non-linear motion and sparse observations, and a life-cycle memory with Confidence Decay Distance and LPF to preserve identities through occlusion. Experiments on KITTI and an indoor office dataset show state-of-the-art performance for LiDAR-based 3D pedestrian tracking, while also highlighting computational trade-offs due to long-term memory; the work points toward future exploration of end-to-end multi-modal, multi-sensor approaches.

Abstract

Person detection and tracking (PDT) has seen significant advancements with 2D camera-based systems in the autonomous vehicle field, leading to widespread adoption of these algorithms. However, growing privacy concerns have recently emerged as a major issue, prompting a shift towards LiDAR-based PDT as a viable alternative. Within this domain, "Tracking-by-Detection" (TBD) has become a prominent methodology. Despite its effectiveness, LiDAR-based PDT has not yet achieved the same level of performance as camera-based PDT. This paper examines key components of the LiDAR-based PDT framework, including detection post-processing, data association, motion modeling, and lifecycle management. Building upon these insights, we introduce SpbTrack, a robust person tracker designed for diverse environments. Our method achieves superior performance on noisy datasets and state-of-the-art results on KITTI Dataset benchmarks and custom office indoor dataset among LiDAR-based trackers.

Figures (5)

• Figure 1: The SpbTracker pipeline associates results from 3D detector predictions with predictions from person motion models. The process involves three main cases: 1. Matched Pairs: Detections and trajectories that are initially matched undergo a two-stage association process. If they remain matched after this process, they are processed through D-UKF (Discrete Unscented Kalman Filter) and LPF (Low-Pass Filter) modules. Pairs with scores exceeding the F1-score threshold are then assigned to either active tracklets or candidate tracklets. 2. Unmatched Trajectories: These are processed through CDD (Confidence Decay and Deletion) modules. Their scores are then adjusted based on their distance from the ego vehicle. Trajectories with scores falling below the death threshold are either deleted or moved to the candidate tracklet pool. 3. Unmatched Detections: Among the unmatched detections, those with scores above the F1-score threshold are added to the birth pool or the candidate tracklet pool.
• Figure 2: The graphs illustrate the relationship between correlation score threshold and two metrics: AMOTP (left) and AMOTA (right) of pedestrian. As the threshold increases, both metrics decrease, indicating a decline in tracking performance. This ablation study was conducted using the DSVT detector model on the KITTI validation dataset. The results highlight the importance of carefully selecting the correlation score threshold to optimize tracking accuracy and precision.
• Figure 3: Comparison of association metrics
• Figure 4: Long-term tracking persistence
• Figure :