A FairMOT approach based on video recognition for real-time automatic incident detection on expressways

An accurate, fast and real-time expressway automatic incident detection (AID) algorithm can not only reduce the burden of expressway management personnel, but also increase the safety and reliability of expressway travel. Aimed at the accuracy from surveillance video, FairMOT is initially transferred from detecting humans to abnormal incidents for expressways, while UA-DETRAC vehicle dataset is employed to train and evaluate YOLOv3 + DeepSORT, YOLOv5 + DeepSORT and JDE. The comparison on evaluation indexes demonstrates that FairMOT improves vehicle tracking effect, and the accuracy is better than the current mainstream algorithms listed above. In the case study, the length change of the track vector is employed to determine whether the car is stopped, and the relationship between the track vector and the center dividing line vector is adopted to decide whether the vehicle is in reverse. The real surveillance video verifies that the proposed FairMOT can detect parking and reversing quickly and accurately. The results can provide an alternative and benefit the automatic incident detection.