Multi-Scale Vehicle Detection and Tracking Method in Highway Scene

Based on tracking-by-detection strategy, this paper proposes a detection and tracking method for multi-scale vehicle targets. The method consists of object detection and trajectory correlation. Targeted at object detection, an improved SNIPER sampling strategy was proposed, which combined with Faster R-CNN deep learning model to achieve variable scale vehicle detection. For the trajectory correlation, motion features, apparent features and trajectory features are matched to achieve the matching between the detection target and the vehicle trajectory. The experimental results show that the method in this paper has made significant improvements over previous methods, both in accuracy and speed, and can detect and track multi-scale vehicle targets in highway scenes in real time.