Research on Binocular Vision Ranging Algorithm Based on Improved YOLOv5
In this paper,the improved YOLOv5 and the SURF algorithm are integrated to address the limited resources issue when deploying on resource-constrained embedded systems or mobile devices due to the high computational complexity and memory requirements of the SURF algorithm. An object detec-tion model based on the YOLOv5_mobilenet backbone network is constructed and the SE attention mechanism and the feature fusion strategy are introduced. On the KITTI dataset,recognition rate rea-ches 88.50％,which is a 0.7％improvement compared to the original YOLOv5. The number of param-eters decreases by 81. 94％,and the floating-point operation decreases by 86.14％. By detecting inter-est regions through object detection,stereo matching and disparity calculation are performed on these regions. Experiment results show that on 500 stereo images,average disparity calculation time for the entire image is 0.574 seconds,while the average calculation time using the SURF algorithm on interest regions is only 0.035 seconds,increasing the computational efficiency by 15 times. Additionally,through multiple experimental calculations,average error rate is determined to be 2.23％.
万方数据
维普
