Lightweight head-based multi-object tracking in crowded built environments

Abstract

Multi-object tracking (MOT) provides an effective approach for occupant monitoring in intelligent built environments. Real-time MOT is challenging on resource-constrained edge devices. This study establishes a systematic evaluation framework for lightweight head-based MOT, examining structured channel pruning and quantization across the complete detection-tracking pipeline. Experiments on Cchead use five detectors and multiple trackers. We comprehensively compare different pruning criteria and pruning strategies, showing that layer-wise pruning with fine-tuning reduces complexity while preserving tracking performance. We further compare post-training quantization (PTQ) and quantization-aware training (QAT) in terms of accuracy preservation, tracking stability, and deployment efficiency; QAT improves AP50:95 by up to 13.33% for PPYOLOE-S. Practical deployment on Jetson Nano, Jetson TX2, and Intel Neural Compute Stick 2 achieves latency reductions exceeding 48%. Robustness tests under Gaussian noise show up to 158.41% relative improvement with QAT. These results characterize trade-offs among efficiency, tracking performance, and resilience for edge MOT deployment.