| With the continuous development of deep learning and big data technology,the robot industry has been developing rapidly.The acquisition of object and distance information is an important task of robot environment perception and an important research direction of computer vision.Target detection based on deep learning has achieved great success in target acquisition,but these models are often large and not easy to deploy on mobile devices,and the accuracy of small model detection is low.At the same time,the matching accuracy of traditional stereo matching algorithms is easily affected by illumination,occlusion,and other factors.To solve these problems,this thesis studies the task of object detection and stereo matching based on binocular vision,and carries out the following work:(1)To solve the problem of low matching accuracy of the traditional stereo matching algorithm under the influence of illumination and other factors,a multi-cost fusion matching cost calculation method is proposed in the cost calculation.The improved gradient cost,color cost,and Census transform are combined with weight to reduce the influence of illumination on matching accuracy.At the same time,the adaptive weight-guided filtering is improved to aggregate the cost space.The improved stereo matching algorithm is evaluated using the Middlebury v3 platform,and the bad4 error matching rates in all regions and non-occluded regions are 19.9% and 13.2%respectively.(2)The YOLOX-s model with fewer model parameters is selected for target detection.Given the low detection accuracy caused by insufficient feature extraction of YOLOX-s,this thesis proposes an improved G-SPPCSPC module to replace the original pyramid pooling layer and uses the attention mechanism to enhance features of different scales.The improved YOLOX-s algorithm was verified in the annotated SODA10 M dataset,and the detection accuracy of m AP was improved by 2.1%.At the same time,the VOC dataset is used to compare with the YOLOv5-s algorithm.The improved algorithm in this thesis is 4.1% higher than YOLOv5-s’ s m AP,which proves the effectiveness of the improved target detection algorithm in this thesis.(3)Assemble the binocular camera,calibrate the camera by the Zhang Zhengyou calibration method,and use the parameters obtained from calibration to correct the image.Combined with the above work,the depth of the target on the corrected image is estimated.Through testing 10 groups of single target images,the average error of the binocular vision scheme proposed in this thesis is 5.06% within 1-5 meters.The experimental results show that the stereo matching algorithm and object detection algorithm proposed in this thesis have high accuracy,and the effectiveness of the proposed algorithm is proved by comparison with other algorithms.At the same time,the experimental results of binocular distance measurement prove that the proposed algorithm has a small error in the target detection and distance measurement task,and can be effectively applied in binocular vision environment perception. |
