IMPROVED FCOS ALGORITHM FOR VEHICLE DETECTION

Aimed at the problems of high error rate and slow detection speed in vehicle detection, an improved fully convolutional one-stage object detection vehicle detection method is proposed. An intersection and union ratio loss function considering multiple geometric factors was introduced, which improved the phenomenon that it was difficult for high aspect ratio vehicles and parallel vehicles to regress accurately in the training process. Multiscale convolution was used to combine multi-dimensional features information, and the robustness of the algorithm to different scale detection was enhanced. According to the scene of vehicle detection, the regression scale was improved to improve the reasoning accuracy of the model. The experimental results show that this method can significantly improve the detection accuracy while maintaining the detection speed in vehicle detection tasks.