Adaptive Frame Sampling for CPU-Based Multi-Camera Person Re-ID

Abstract

Person re-identification (Re-ID) in multi-camera surveillance is typically designed for GPU hardware, limiting deployment in the CPU-only environments common across resource-constrained regions. We propose an adaptive frame sampling framework for multi-camera person Re-ID that operates in real-time on CPU-only systems while preserving identity tracking accuracy. The proposed system integrates a motion-triggered adaptive sampling mechanism with a YOLOv8n person detection model and a lightweight CNN-based Re-ID module. Rather than executing deep learning inference on every captured frame, the system selectively triggers inference when a frame differencing-based motion intensity score exceeds a defined threshold, skipping redundant frames during static or near-static scenes. The detection model was trained on the UCSD Pedestrian 1 dataset over 50 epochs, achieving a final mAP50 of 0.979 and a peak F1 score of 0.95. The Re-ID module extracts 256-dimensional L2-normalised appearance embeddings from detected person crops and performs identity assignment via cosine similarity matching against a rolling gallery. Comparative evaluation across three independent video experiments demonstrated that the adaptive sampling strategy reduced processed inference frames by an average of 89.3% and total processing time by 71.1%, while achieving an effective frame rate of up to 56.5 FPS on CPU hardware, with 100% identity consistency maintained across all conditions. These findings demonstrate that high-accuracy multi-camera person Re-ID is achievable at a fraction of the computational cost of conventional approaches, making real-time surveillance feasible in low-resource settings and expanding access to advanced video analytics beyond GPU-equipped deployments.