Research On Binocular Vision Object Decteion And Localization Based On Deep Learning

With the continuous development of artificial intelligence technology,how to get more abundant information through machine has become a key research.Object detection and positioning technology have become a hot issue in the field of artificial intelligence,which has a wide application prospect and a very huge market value in many fields.It can be used in the fields of automatic driving assistance,UAV,virtual interaction and military reconnaissance.In this paper,the object detection algorithm and binocular stereo matching algorithm are studied deeply.The target detection and location functions are realized by combining the deep learning target detection algorithm and stereo matching algorithm,and the existing problems are solved.Existing object detection algorithms based deep learning,while achieving the task of target detection quickly and accurately,need to consume a large number of computing resources,and the amount of computing model parameters is large,which makes the deep learning-based target detection algorithms Facing huge challenges in future applications.In view of this,based on the network of YOLO v3 object detection algorithm based on regression thinking,combined with Mobile Net v2 lightweight network for feature extraction,this paper proposes a YOLO v3 network structure based on inverted residuals;and for deep separability Convolutional feature extraction is relatively sparse,leading to low prediction accuracy.Combining deep separable convolution and residual structure to design a prediction branch network effectively improves the accuracy of target detection.At the same time,considering the different effects of the target size in the image on the loss function in the process of target prediction,a variable-scale weighted GIo U loss function was designed,and the weight of the candidate window was calculated based on the predicted candidate window size.The experiments show that the amount of network structure parameters is reduced and the operating speed is significantly improved after the improvement,while maintaining good detection accuracy.The stereo matching algorithm includes local stereo matching algorithm and global stereo matching algorithm.The local stereo matching algorithm has two cost aggregation methods,which are adaptive support window and adaptive support weight.The adaptive window relies on the color threshold as the judgment basis,so that the window of the intersection region of pixels at the edge will contain more irrelevant regions,which will cause interference to the subsequent parallax value.In this paper,an edge constrained adaptive guided filtering local stereo matching algorithm is proposed,which combines the guided graph filtering with the cost aggregation algorithm of the adaptive region to construct the adaptive guided filtering,so as to ensure that there are more effective windows in the cost aggregation process,make full use of the differences of local features,and use the linear threshold method to calculate the color threshold.At the same time,the maximum extension threshold of arm length is set at the boundary point to make the boundary point contain more effective pixel information,which is conducive to edge protection of disparity map.In this paper,the object detection algorithm of deep learning and stereo matching algorithm are combined to realize the target detection by the binocular camera,and the relative pose of the target and the camera is calculated,thereby achieving the target detection and positioning function.Furthermore,the combination of the object detection algorithm and stereo matching algorithm proposed in this paper is used to analyze indoor positioning experiments.