Vehicle Detection And Spatial Location In Video Badsed On Deep Learning

In recent years,due to the continuous popularization of artificial intelligence,intelligent vehicles and vehicle assisted driving have gradually become the direction of technological change in a new round,and achieved quite good results.Among them,the detection,recognition,segmentation and ranging of objects(including pedestrians,cars and traffic lights)in the video stream is an important part of the intelligent vehicle assistant driving system.Since Alex Net won the first place in the Image Net competition in 2012,in-depth learning has made great achievements in the fields of target detection,recognition and segmentation.At the same time,in the field of computer vision,using binocular stereo vision to detect the distance of the target has gradually attracted the attention of many researchers.However,the above scheme still has the problem of insufficient real-time performance caused by too much computation.Target detection,recognition and segmentation are the premise and basis of target ranging.Because the purpose of this paper is to obtain the accurate segmentation image of the target,and the current algorithm that combines detection,recognition and segmentation is only Mask R-CNN,so the basic scheme of target detection,recognition and segmentation in this paper is Mask R-CNN algorithm;at the same time,using binocular stereo vision to distance the target.And in view of the above several algorithms involved in the lack of real-time,the following solutions are proposed.First of all,aiming at the problem of too many parameters,too much computation and lack of real-time performance of Mask R-CNN,this paper proposes the algorithm of Mask R-CNN based on Mobile Net V2,which uses Mobile Net V2 to replace the feature extraction network in Mask R-CNN,effectively reduces the operation time of Mask R-CNN.Finally,the measured data,including the average running time and the m AP(mean Average Precision),are given to verify the validity.Secondly,aiming at the problem that the Mask R-CNN algorithm based on Mobile Net V2 can not achieve real-time performance,this paper further proposes the Mask R-CNN algorithm based on Mobile Net V2 combined with optical flow method,which uses optical flow method to track the identified target in consecutive frames,effectively reducing the running time of the Mask R-CNN algorithm based on Mobile Net V2.Finally,on the problem of extracting the depth of the target,this paper first summarizes the binocular stereo vision algorithm,and in view of the lack of real-time performance of the matching algorithm in binocular stereo vision,proposes a binocular matching algorithm based on the Mobile Net V2 Mask R-CNN algorithm.The innovation of the algorithm is that it directly uses the target information obtained by the Mask R-CNN algorithm based on Mobile Net V2,avoids the time loss caused by the full picture matching,and gives the ranging error rate to verify the algorithm.