Fast and accurate obstacle detection of manipulator in complex human–machine interaction workspace

Abstract

Abstract In order to solve the problems of a large amount of calculation, poor real-time performance, and insufficient detection accuracy in an existing robot obstacle detection system, this paper proposes an improved lightweight You Only Look Once Version 3 algorithm for obstacle detection and tracking, by combining the DeepSort algorithm with a three-dimensional model velocity estimation method. First, the depthwise separable convolution and convolution kernel pruning methods are used to lighten the network. Second, to meet the requirement of dynamic obstacle detection, the DeepSort method is used to track dynamic obstacles and estimate their moving speed. Third, a method of bounding-sphere modeling and discrete marking is proposed to simplify the obstacle model. The experimental results show that: the model size is reduced by 92.11%, which suggests that the proposed lightweight detection system is more suitable for industrial applications; the detection speed in the actual scene reaches 118 frames per second and the average detection accuracy is 92.6%, while the multiple objects tracking accuracy and multiple objects tracking precision are all above 0.8 in the dynamic obstacle tracking process. Static and dynamic obstacles in the workspace can be detected simultaneously with an average positioning accuracy error of 1.5 cm. The proposed model can not only effectively improve the detection efficiency of obstacles in the manipulator workspace but also enhance the safety of industrial robot systems.