Abstract
DensePeds is a novel pedestrian tracking algorithm optimized for high-density crowds (>2 people/m²). It integrates a Front-RVO (FRVO) motion model, which uses elliptical approximations for front-facing cameras, with sparse feature extraction via Mask R-CNN segmentation. This approach achieves state-of-the-art accuracy in dense environments while operating up to 4.5x faster than previous leaders.
Introduction
Pedestrian tracking is a cornerstone of autonomous navigation and urban surveillance, but performance typically degrades in dense crowds. Standard "tracking-by-detection" systems rely on rectangular bounding boxes. In packed environments, these boxes overlap significantly, causing the appearance features of one individual to be corrupted by noise from neighbors and the background. This often leads to identity switches or the complete loss of tracks.
Furthermore, traditional motion models like Constant Velocity or standard Reciprocal Velocity Obstacles (RVO) often fail in front-facing or elevated camera views. RVO typically assumes pedestrians are circles viewed from a top-down perspective. In front-facing views, this circular assumption creates "false overlaps," leading algorithms to predict non-existent collisions and hindering accurate path prediction.
Methodology
The DensePeds framework introduces two primary innovations to overcome the limitations of traditional tracking in dense environments.
Front-RVO (FRVO) Motion Modeling
FRVO is a non-linear motion model that extends the RVO framework to accommodate front-facing and elevated camera perspectives. Instead of modeling pedestrians as circles, FRVO utilizes elliptical approximations where the axes correspond to the height of the face and the width of the shoulders. This geometric shift allows the algorithm to more accurately predict pedestrian dynamics by considering collision avoidance constraints specific to the camera's perspective, effectively eliminating the error margin that grows as people get closer in circular models.
Sparse Feature Extraction
To mitigate the noise inherent in overlapping bounding boxes, DensePeds employs Mask R-CNN for pixel-wise segmentation. Instead of extracting features from a noisy rectangular region, the system "zeros out" the background and adjacent pedestrians. This process creates a "sparse" feature vector that uniquely represents the individual being tracked, acting as a spatial filter to prevent the corruption of a person's appearance memory.
The system concludes its pipeline by using the Hungarian algorithm to match predicted movements and cleaned features across video frames.
Experimental Evaluation
The researchers evaluated DensePeds using both standard benchmarks and a new, more challenging dataset:
- MOT15 & MOT16: Standard multi-object tracking benchmarks with relatively sparse densities (~0.2 persons/m²).
- New Dense Crowd Dataset: A collection of 8 high-density videos (2.0 to 2.7 persons/m²) captured from elevated stationary cameras in public areas, providing a realistic testbed for metropolitan-level crowding.
Key Results
- Efficiency: DensePeds is 4.5x faster than prior state-of-the-art algorithms on the MOT benchmark due to its optimized search space and linear programming for collision avoidance.
- Accuracy in Density: On the new dense crowd dataset, DensePeds improved Multi-Object Tracking Accuracy (MOTA) by 2.6% and reduced false negatives by 11% compared to leading methods like MOTDT and MDP.
- Ablation Analysis: Replacing FRVO with a standard constant velocity model caused the "mostly lost" (ML) track metric to jump to nearly 100%, confirming that the motion model is critical for performance in dense environments.
Conclusion
DensePeds demonstrates that fundamental geometric fixes and precise spatial filtering can significantly improve tracking performance in challenging environments. While the benefits are most pronounced in dense crowds—where reciprocal velocity avoidance is most active—the system remains competitive on sparse benchmarks, ranking in the top 13% of published methods on MOT15. Future work will focus on expanding the dense crowd dataset and potentially integrating more complex behavioral models beyond simple collision avoidance.
Summary updated