| Badminton is an indoor sport for people of all ages,popular all over the world for its ability to increase social interaction and exercise.In the process of badminton practice,a large number of badminton balls will be scattered on the ground,if the manual picking method is adopted,not only increase the investment of human resources,but also the pickup efficiency is not high.With the continuous improvement of the intelligence of badminton robots,they have more and more comprehensive functions,and can carry out a variety of different actions such as hitting,picking,and recycling badminton.In the process of picking up the badminton,it is necessary to classify and detect and measure the badminton to meet the robot’s perception and positioning of the ground badminton.This paper focuses on the key algorithms of indoor ground badminton classification detection and ranging,and proposes a ground badminton classification detection and ranging method based on binocular vision:1.The imaging model and coordinate system of binocular camera are constructed,the internal and external parameters of binocular camera are obtained,the correspondence between real-world targets and image targets is established,and the error caused by lens distortion of the camera is eliminated by the stereo correction algorithm.Indoor ground badminton images were collected,defective badminton and complete badminton were marked,indoor ground badminton dataset was made,and the traditional target detection algorithm was simulated by using the dataset to study the process of feature extraction by convolutional neural network.2.Improve the key technology of distance measurement algorithm,stereo matching,and propose a fast stereo matching algorithm based on the improved Census transform.Using the average of eight equidistant pixel points around the center pixel as the reference value,reduce the dependence on the center pixel and improve the robustness of the algorithm;Utilizing Sobel operator to enhance the features of target edge information,reducing aggregation paths in cost aggregation,and improving algorithm execution efficiency.3.In response to the difficulty of detecting small and easily missed badminton targets,a small target detection network YOLOv5 SGCM algorithm is proposed.The Mosaic algorithm is used to enhance the generalization ability of the dataset,add a small target detection layer P2 and integrate CBAM attention mechanism.Ghost Net is used to reduce redundancy in feature maps.Comparative and ablation experiments are designed to calculate the detection accuracy of binocular cameras at different distances,and the algorithm is deployed on the hardware platform Jetson Xavier nx,Real time classification and ranging effect detection of badminton. |
