Research On Tracking Of Coding Target On Linear Translation Using Deep Learning

Computer vision is an important research field in deep learning and target tracking is a key technology in computer vision.There are many applications of target tracking,such as automatic driving and intelligent monitoring.The essence of target tracking is to obtain real-time locations of target objects.Target detection algorithms generally need to be effectively combined with various image algorithms,such as convolutional neural networks based on region-selective search algorithms(R-CNN),which is the current mainstream solution and the core technology in the target tracking framework.With the improvement of target detection technology,the target tracking technology has gradually developed to the height of semantic segmentation.In this thesis,target tracking algorithm is applied to the high-precision displacement measurement of the target.The measurement target is placed on the motorized translation.The positions of target are hybrid coded,including absolute code channel and incremental code channel.The absolute code channel is used to locate the initial position.The track is used to obtain the fine position.The convolutional neural network is used to build an end-to-end target tracking framework,and the real-time positioning of feature targets is realized by detecting the coding information at a specific location.In order to achieve sub-pixel precision subdivision of the target location,a fully connected neural network is used to build an accurate subdivision network.In the end-to-end convolutional neural network,spatial pyramid pooling is introduced in the last pooling layer to deal with the problem of different sizes of input images.In addition,the final result classification is different from the SVM classification used by traditional R-CNN.In this thesis,94.0% accuracy is achieved by using the fully connected layer.Finally,the barcode sequence on the absolute code channel is segmented according to the positioning information in the incremental code channel.For the segmented barcode,the recognition accuracy of 96.6% can be achieved.Experimentally,by adding a macro lens to the rear camera of the mobile phone to obtain the target coded picture,the system realized the recognition accuracy of 2.6 through the trained neural network recognition.The system and device have portable,high-accuracy measurement capabilities,and the acquisition accuracy can be further improved by increasing image acqusition.