Dominik Kroupa
Supervisor(s): prof. Ing. Adam Herout Ph.D.
Brno University of Technology
Abstract: Multi-target multi-camera pedestrian tracking (MTMCT) plays a crucial role in urban surveillance, public safety, and crowd behavior analysis, contributing to the advancement of smart city infrastructure by providing analysis of pedestrian movement through given areas. However, due to challenges such as the presence of severe occlusion and changes in appearance throughout the scene, robust and accurate MTMCT remains a significant challenge in complex environments. To address these challenges, this study proposes an offline pipeline for pedestrian tracking, consisting of three main stages: (1) generation of single-camera tracklets through pedestrian detection (with keypoint estimation) and appearance feature extraction, (2) refinement and completion of tracklets using appearance features and strategies to reduce identity switches, and (3) inter-camera association (ICA) via global ID assignment leveraging appearance features. Additionally, a model trained on detected body keypoints is employed for ground position estimation. The solution was evaluated in the AI City Challenge MTMCT Track in the previous year, achieving a Higher Order Tracking Accuracy (HOTA) score of 31.52%.
Keywords: Computer Vision
Full text: Year: 2025
