Research On Target Detection And Ranging Based On Binocular Vision

The detection and ranging of front targets of construction vehicles which is based on machine vision has become a hot issue in current research.It aims to identify the front targets of construction vehicles and measure their distance,so that construction vehicles can perform safe and autonomous operations.At present,a lot of research has been carried out domestically and abroad,they are based on intelligent technology related to mechanical equipment,and many research results have been achieved.However,the real-time performance of the detection technology and the accuracy of the ranging technology still need to be improved.The main research contents of this article are summarized as follows:(1)This paper studies the Yolov3 network target detection algorithm.In view of the problem of long detection time of the network,this paper reduces the convolutional layer of the Yolov3 network and the prediction output by one layer,and then changes the confidence value of the Yolov3 network to reduce the recognition of high similarity but unrelated features to the recognized object.And ultimately reduces the detection time by 85%.(2)To solve the problem of the Yolov3 network which has poor target detection effect and localization effect on dense small targets,this paper studies the Faster-RCNN network target detection algorithm.Faster-RCNN uses the ROI Pooling layer.When feature extraction is performed on the original image,the corresponding feature map will be extracted,then the corresponding ROI will be mapped,as to obtain a more accurate candidate frame,instead of directly converting the whole as Yolov3.The image is divided into multiple grids.Therefore,the positioning accuracy of Faster-RCNN is better than Yolov3,and its detection accuracy is improved by 3.6% compared with Yolov3 network(3)Performed some research the binocular camera ranging algorithm for the problem that the binocular camera has distortion when collecting images,the Zhang's calibration method is used to calibrate the binocular camera to obtain the internal and external parameters of the binocular camera.To solve theproblem of high complexity and poor real-time performance of the global image matching algorithm,a local image block matching algorithm is used to perform stereo matching on the left and right images.Through experimental comparison,the local image block matching algorithm has good real-time performance and high matching accuracy.Using the obtained depth map,the world coordinate system z-axis size and the actual distance of the pixel point from the camera in the real world can be obtained,that is,distance information can be generated.(4)In the case of ordinary binocular cameras at night or insufficient light,the collected images are not good,at the same time,when there is a solid color area in the image collected by the binocular camera,these are not conducive to subsequent image matching,which will affect its ranging.For the problem of accuracy,this paper applies the method of structured light to the ranging process of binocular cameras.When collecting images,the camera emits structured light to generate some light spots with textured areas on the surface of the object,which improves the image matching rate and improves the accuracy of distance measurement in low-light environments.At last,combining the Faster-RCNN network target detection algorithm with the binocular vision ranging algorithm to prove the effectiveness and feasibility of this method in target detection and ranging of construction vehicles are verified.