Target Detection and Tracking of Ground Mobile Robot Based on Improved Single Shot Multibox Detector Network

Abstract

To solve the problems of slow labeling speed of the traditional labellmg data set establishment method, and slow running speed of target classification and detection algorithm based on Single Shot Multibox Detector(SSD) deep learning network, this paper proposes a fast data set labeling algorithm and a fast SSD network for target real-time detection and tracking research. First, a data set is established quickly by using TLD target detection and tracking algorithms, cropping and mirroring methods are used to strengthen the data set. Then, SSD backbone network is improved based on depth-wise separable convolution to establish a fast SSD network. Finally, the ground mobile robot in RoboMasters(RM) competition is used as the detection and tracking target indoors and outdoors, as well as with other different scenarios with shield to test the real-time performance, accuracy and effectiveness of the algorithm. The results show that compared with traditional SSD network research, in terms of the analysis and processing system deployed on low-performance hardware, the improved fast SSD network can better meet the real-time requirements of target detection and tracking.