OccluTrack: Rethinking Awareness of Occlusion for Enhancing Multiple Pedestrian Tracking

TL;DR

This work proposes an adaptive occlusion-aware multiple pedestrian tracker, OccluTrack, that outperforms state-of-the-art methods on MOTChallenge and DanceTrack datasets and develops a pose-guided re-identification module to extract discriminative part features for partially occluded pedestrians.

Abstract

Multiple pedestrian tracking is crucial for enhancing safety and efficiency in intelligent transport and autonomous driving systems by predicting movements and enabling adaptive decision-making in dynamic environments. It optimizes traffic flow, facilitates human interaction, and ensures compliance with regulations. However, it faces the challenge of tracking pedestrians in the presence of occlusion. Existing methods overlook effects caused by abnormal detections during partial occlusion. Subsequently, these abnormal detections can lead to inaccurate motion estimation, unreliable appearance features, and unfair association. To address these issues, we propose an adaptive occlusion-aware multiple pedestrian tracker, OccluTrack, to mitigate the effects caused by partial occlusion. Specifically, we first introduce a plug-and-play abnormal motion suppression mechanism into the Kalman Filter to adaptively detect and suppress outlier motions caused by partial occlusion. Second, we develop a pose-guided re-identification (Re-ID) module to extract discriminative part features for partially occluded pedestrians. Last, we develop a new occlusion-aware association method towards fair Intersection over Union (IoU) and appearance embedding distance measurement for occluded pedestrians. Extensive evaluation results demonstrate that our method outperforms state-of-the-art methods on MOTChallenge and DanceTrack datasets. Particularly, the performance improvements on IDF1 and ID Switches, as well as visualized results, demonstrate the effectiveness of our method in multiple pedestrian tracking.

Figures (8)

• Figure 1: The motivation of our proposed OccluTrack. (a) Left: the SOTA tracker botsort struggles with abnormal motions caused by partial occlusion, leading to inaccurate motion estimation. Middle: They also lack pose guidance during partial occlusion, resulting in unreliable appearance features. Right: Treating occluded and non-occluded pedestrians equally in distance measurement hinders optimal association. (b) In contrast, OccluTrack incorporates abnormal motion suppression (left), pose-guided Re-ID (middle), and occlusion-aware distance measurement (right) to address these limitations.
• Figure 2: Overview of our proposed tracker. First, we utilize YOLOX yolox as the object detector to obtain bounding boxes and separate them into high-score and low-score detections. For high-score detections, the pose-guided Re-ID module in (a) is used to extract appearance features. Second, the abnormal motion suppression Kalman Filter (AMS KF) is proposed to obtain $Predictions_{k-1}$ from previous $tracklets_{k-1}$. Based on the predictions, the association module with occlusion-aware distance measurement in (b) first associates the $Predictions_{k-1}$ of previous frames with the high-score detections (with feature embeddings), and then the second association matches unmatched predictions with low-score detections. Finally, the abnormal motion suppression Kalman Filter in (c) is updated based on the matching results obtained from the association step. Specifically, the proposed AMS KF first detects the abnormal detections with abnormal changes. The $tracklets_{k}$ and parameters of the KF are updated correspondingly by considering the outlier detections caused by partial occlusion.
• Figure 3: Architecture of our proposed pose-guided Re-ID module. It first extracts appearance features from the cropped image and estimates the pose of the occluded person. After that, body-part features are generated by applying body-part heatmaps to appearance features. By utilizing our proposed local and global necks, local and global feature embeddings are obtained for training. With the adaptive fusion, local and global embeddings are combined for inference.
• Figure 4: Combination of body keypoints to form different body parts. The keypoints are grouped into six body parts according to the given rules.
• Figure 5: Ablation study of abnormal motion suppression parameter $\alpha_0$ on the MOT17 validation set. The $\alpha_0$ is adjusted from 0 to 1. IDF1 and MOTA jointly reach the peak when setting $\alpha_0$=0.2.