Dynamic Target Detection and Tracking in Water for Mobile Robot Based on Deep Learning

Abstract

Digital image processing is an important part of information perception of mobile robots. Unlike simple geometry, the shape of real objects is always irregular. An accurate and real-time dynamic image processing strategy for mobile robots based on deep learning is developed in this paper. To ensure the dynamic target in water always in the center of the mobile robot’s vision field, a target object detection strategy is designed according to the YOLO-V3 algorithm. 2279 pictures of the target object at different draught depths and different motion directions are collected to retrain the YOLO-V3 model. After testing, the accuracy of the model reaches 94.82%. Besides, considering the high accuracy and high efficiency of the Siamese network, SiamFC (a highly representative algorithm) is selected to support dynamic target tracking. An improved target tracking algorithm based on detection and supervision feedback is designed based on IOU (Intersection over Union) concept. Also, to guarantee the smooth motion of the mobile robot, a strategy of terrain information perception and obstacle terrain passing based on lidar scanning is designed. By analyzing the results of lidar scanning, the mobile robot can judge and avoid the obstacle terrains. The real-time and accuracy of each algorithm is verified by a comprehensive experiment of dynamic target search, detection, and tracking.