Pedestrian tracking using probability fields and a movement feature space

Negri, Pablo; Garayalde, Damian · 2017 · DOAJ

DOI: 10.15446/dyna.v84n200.57028

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper presents a pedestrian tracking framework that utilizes a Movement Feature Space (MFS) and probability fields to improve tracking accuracy in complex, real-world video sequences. The authors address the challenges of pedestrian tracking, such as changing appearances, non-rigid structures, and occlusions, by proposing a method that outperforms existing algorithms with lower computational complexity. The approach is evaluated on two public datasets, PETS2009 and TownCentre. The methodology centers on a "Target Framework" where each pedestrian is modeled as an autonomous entity tracked via a state machine. The system operates within a Tracking-by-Detection pipeline. First, the MFS is used to extract motion features, specifically histograms of oriented level lines, which serve as descriptors for a pedestrian detector based on boosted classifiers. This detector generates regions of interest (ROIs) and confidence scores. To track targets, the system constructs two probability fields: a Detection Field, derived from detector output scores, and an Appearance Field, generated by analyzing corners in the MFS. The Appearance Field is particularly robust as it ignores static background clutter. An iterative tracking algorithm, inspired by Mean Shift and Lucas-Kanade methods, uses these fields to estimate displacement vectors by maximizing the Bhattacharyya coefficient between distributions in consecutive frames. A pyramidal search strategy ensures fast convergence. Each target is managed by a state machine with states including INIT, STILL, WALKING, VERIFY, and END. The VERIFY state is activated when a detection is lost, allowing the system to continue tracking using the Appearance Field. An offline pruning stage refines the results by filtering false positives—identified by targets that remain only in INIT and VERIFY states—and concatenating broken trajectories. The concatenation process employs a Kalman filter to estimate trajectories and matches lost targets with new detections using Mahalanobis distance metrics. The results demonstrate that this approach achieves superior tracking performance compared to state-of-the-art systems reported in the literature, despite having lower computational complexity. The use of the MFS allows the system to operate effectively in scenes with cluttered backgrounds and varying illumination, as the motion-based features are robust to static noise. The framework successfully handles partial information and detector failures, providing a reliable solution for pedestrian tracking in uncontrolled environments.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• pedestrian behavior perception