Lightweight Object Detection Method for Mobile Robot Platform

Abstract

∗We present some experienced improvements to YOLOv5s for mobile robot platforms that occupy many system resources and are difficult to meet the requirements of the actual application. Firstly, the FPN + PAN structure is redesigned to replace this complex structure into a dilated residual module with fewer parameters and calculations. The dilated residual module is composed of dilated residual blocks with different dilation rates stacked together. Secondly, we switch some convolution modules to improved Ghost modules in the backbone. The improved Ghost modules concatenate feature maps obtained by convolution with ones generated by a linear transformation. Then, the two parts of feature maps are shuffled to boost information fusion. The model is trained on the COCO dataset. In this paper, mAP_0.5 is 56.1%, mAP_0.5:0.95 is 35.7%, and the speed is 6.1% faster than YOLOv5s. The experimental results show that the method can further increase the inference speed to ensure detection accuracy. It can well solve the task of object detection on the mobile robot platform.