Intersection vehicle detection technology based on DeepSORT algorithm

This paper proposes a real-time multi-object tracking algorithm based on improved YOLOv7 and DeepSORT to address the issues of poor tracking performance and target identity switch caused by occlusion in the current intersection detector's field of view. To enhance the perception capability of the network towards the targets and improve the accuracy of vehicle detection, an attention mechanism is added to the backbone network of YOLOv7. A constant acceleration Kalman filter is used to optimize the vehicle motion model for the tracking part. DenseNet-121 is employed as the backbone network for feature extraction replacing ResNet. This substitution enhances the network's ability to extract vehicle features. The algorithm is evaluated using video data captured from actual road intersection detector perspectives as the test set. Compared to the original algorithm, the proposed algorithm achieves an average accuracy improvement of 6 percentage points and effectively resolves the problem of vehicle re-identification after occlusion.