YOLO-LIGHT: A Real-Time Lightweight Model Based on Limited Edge Computing for Traffic Scene Detection and Tracking

Intelligent transportation systems (ITSs) play a vital role in addressing urban traffic safety challenges. The deployment of lightweight and efficient algorithms for conducting vehicle detection and tracking on edge devices is a key requirement, due to the limited computational resources of these devices. To address this challenge, we propose YOLO-LIGHT, a real-time vehicle detection model, and Ve-Track, a lightweight tracking algorithm. YOLO-LIGHT improves the feature extraction process and reduces the incurred computational cost. It integrates a novel FastPConv convolution in its detection head, while the DySample and SCAM modules in the neck provide enhanced feature representations. Soft-NMS and AMA_SPPF further strengthen the ability of the model to capture fine details. Additionally, global channel sparsity and pruning are employed to compress the parameters of the model while preserving its detection accuracy. For tracking purposes, Ve-Track builds on ByteTrack by filtering background detection boxes using a “life value” indicator and recalculating scores. An enhanced Kalman filter (EV-KF) is also introduced to better model nonlinear vehicle motions. Experimental results obtained on the modified VisDrone dataset show that YOLO-LIGHT improves the mAP95 metric by 17.6% and exhibits a computational complexity reduction of 82.6%, achieving 51 FPS on an embedded platform. Ve-Track outperforms ByteTrack, with a 6.2% MOTA increase and a 17.2% FPS increase, demonstrating its suitability for real-time ITS applications.