Resource-Efficient Multi-Camera People Tracking : A Comparative Study of Backbones and Frame-Skipping

Sammanfattning

Multi-Target Multi-Camera (MTMC) tracking systems provide significant spatial context for real-time monitoring across disjoint camera edges. With real-time detection and tracking, an MTMC system has the potential to produce actionable insights through deep learning networks. Person Re-identification (Re-ID) can help realize this by enabling coherent identity matching across the camera edges. However, the combined computational demands of detection, tracking, and feature extraction pose significant challenges for scalability. Therefore, this thesis investigates open-world MTMC tracking efficiency by evaluating various Re-ID backbones and fixed frame-skipping to balance inference speed with tracking accuracy and identity consistency in open-world MTMC systems. The proposed framework uses YOLOv11 for detection, ByteTrack for tracking, and compares the performance of OSNet, ResNet-50, and MobileNetV2 for appearance feature extraction. To ensure privacy, the system uses edge-based processing with no transfer of raw video data and regularly removes appearance-specific data of individuals. Experimental results indicate that skipping every other frame increases the system’s relative inference speed by 72 % with negligible loss in tracking accuracy. Furthermore, frame-skipping was found to mitigate noise and prevent the accumulation of degraded embeddings in challenging environments. While OSNet proved to be the superior backbone for Re-ID, the choice of backbone had a limited impact on overall inference speed, confirming the detection and tracking frameworks as the primary system bottlenecks.