ST-DeM: Enhancing DeepSORT with Swin Transformer and Attention-Augmented MobileNetV3 for Robust Illegal Parking Detection

Illegal parking poses significant challenges to urban traffic safety and efficiency. To address the limitations of existing detection methods, such as low accuracy under occlusion and scale variation, this paper proposes ST-DeM, an enhanced illegal parking detection model. Built upon the DeepSORT framework, ST-DeM integrates a Swin Transformer detector for improved multi-scale object detection and a novel CA-MobileNetv3 feature extractor. The latter uses a Coordinate Attention (CA) mechanism to learn robust appearance features, effectively reducing identity (ID) switches during occlusions. A violation is identified if a tracked vehicle remains stationary within a prohibited zone, as determined by centroid displacement and velocity thresholds. Experiments show that ST-DeM achieves a Multi-Object Tracking Accuracy (MOTA) of 64.3% on the MOT16 benchmark, outperforming the DeepSORT baseline while significantly reducing ID switches. On a custom dataset, it achieves a 91.1% detection rate for illegal parking events. This work provides an effective and robust solution for intelligent traffic management.