YOLO-based Multiple Target Detection And Measurement For Robots In Complex Scenes

Multi-target detection and orientation measurement in complex scenes are the key functions in an intelligent robot system.The robot needs to detect and recognize the types of items in the surrounding environment in real time,and obtain the distance and coordinates of each item,so as to lay the foundation for it to respond appropriately to different items and perform motion planning.This paper mainly proposes two lightweight YOLO algorithms that can run in real time on embedded systems,designs and implements a binocular vision multi-target detection and measurement system based on YOLO,and completes the packaging of the algorithm in the robot operating system(ROS).In order to improve the target detection ability of robots in a complex unstructured environment,the thesis uses deep learning technology to build a robot target detection system.Considering the load capacity and computing resources of mobile robots comprehensively,by comparing a variety of deep learning-based target detection models,we finally choose to improve on the basis of YOLOv3 algorithm with better real-time and accuracy,so that it can be embedded in the robot.Quickly detect targets on the system.In this paper,based on the existing simplified version of YOLOv3-tiny,two lightweight solutions are proposed,including YOLO-mini based on "pruning" experiments and MYOLO based on deep separable convolution.After testing the speed and accuracy of the algorithm on the embedded device Jetson TX2,YOLO-mini is used as the target detection network of the system.After obtaining the target area in the image,the target pose measurement based on the binocular camera is realized,the depth map is calculated,and the distance information of each target is obtained,thereby obtaining the three-dimensional coordinates of the target in the camera coordinate system.Furthermore,the target detection and binocular ranging arepackaged into different ROS nodes respectively to complete the deployment on Jetson TX2,which is convenient to integrate with robot navigation,positioning,grabbing and other functions,and promote the engineering application of the algorithm.In the experimental test of the detection system,the indoor object types are selected from the open source data set COCO to form a new data set,the improved target detection model is trained,and its average detection accuracy(mAP,45.9%)is calculated,using different distances under the square The box performs binocular ranging and evaluates the ranging error(the maximum error within a distance of 1 to 3meters is less than 2%).Experiments show that the accuracy and running speed of the multi-target detection and measurement system of the robot meet the actual use requirements of the mobile robot.