Research On The Method Of Object Detection And Tracking Based On Deep Learning

Object detection and tracking is one of the most important researchful hotspots in the field of computer vision.It has been extensively used in many fields such as military warfare,intelligent human-computer interaction,medical imaging diagnosis,etc.Although many classic algorithms concerning the object detection and tracking have been proposed by researchers,but because of the complex variety and uncertainty of real scenes in real life,the research of object detection and tracking still has great challenges.Based on the principle of traditional object detection and tracking method and the practical application of deep learning model in the field of object detection and tracking,the research of the object detection and tracking method based on deep learning has been carried out.The main research work of this thesis consists of the following two parts:(1)In this paper,it makes a deeply study of the object detection model YOLO and analyzes the advantages and disadvantages of the model.Then,the YOLO model is optimized by multi-feature fusion,anchor frame and multi-scale training,which effectively improves the accuracy and efficiency of the object detection.1.Aiming at the problem that the YOLO model has low utilization rate about output characteristic information of the shallow convolution,it is proposed to add a fusion layer based on the original model to realize the fusion of the output characteristic information of different convolutional layers in the network model,and take full advantage of the shallow layer which includes rich information,to improve the ability of detecting small object objects by the YOLO model;2.For the problem of predicting the bounding box of YOLO model,this paper draws on the method of generating the anchor frame in FasterR-CNN,and convolves the feature map to realize and predict the bounding box along with confidence score of each position of image,to improve the accuracy rate of positioning specific objects.3.For the problem that the size of input image about YOLO model is too singleduring the training process,the multi-scale training method is adopted to improve the robustness of the network model.(2)In the process of object tracking,the particle filter tracking model has limitations in selecting features manually.In this paper,a particle filter tracking method based on convolutional neural network is been proposed.The method combines the convolutional neural network model with the particle filter framework,extracting the features of the tracking object through a small-scale convolutional neural network,predicting and estimating the tracking objects in the framework of the particle filtering model.The accuracy of object tracking can be improved effectively by this method.