An Improved Camera-Based People Counting System with Identity Tracking and Multi-state Line-Crossing Logic

Camera-based people counting plays an important role in indoor occupancy monitoring, facility planning, and safety management. However, achieving reliable counting remains challenging in scenes with occlusion, crowded pedestrian flow, and inconsistent tracking identities. This work presents a people-counting framework that integrates deep-learning-based object detection, identity-preserving tracking, and a line-crossing counting logic with optional multi-state validation. The system supports two deployment viewpoints: a bird’s-eye view for head-top detection and an overhead corridor view for full-body tracking. YOLOv5, YOLOv8, and MobileNetV2 SSD detectors were evaluated in combination with SORT, ByteTrack, and DeepSORT tracking algorithms. A three-state trajectory validation strategy is introduced to confirm before-on-after transitions across a virtual counting line, reducing false counts in scenes with motion jitter or partial occlusion. Experimental results indicate that MobileNetV2 SSD with DeepSORT and the three-state logic provides stable counting in the bird’s-eye configuration, while the one-state method remains suitable for corridor scenarios with consistent motion. Overall, the framework offers a practical and flexible solution that can be adapted to different spatial layouts and environmental conditions.