The Research Of Object Detection Methods Based On Deep Learning

As a country with a large population,with the improvement of people's living standard,the increase of vehicles has become more and more obvious,which has brought about traffic safety hazards.Therefore,it is necessary to monitor the vehicles on the road.For human vision,we can easily identify the types and positions of objects contained in an image,but for a computer,it is very challenging to detect and recognize the objects and positions in an image.This paper takes the vehicle image as the research object to detect the vehicle position information in the image,so as to achieve the purpose of monitoring the vehicle in the image.The traditional object detection methods can be divided into four steps: image preprocessing,object region selection,feature extraction and classifier classification.For the selection of object area,the traditional method is the region selection strategy through sliding window,which is of high time and space complexity.In the aspect of feature extraction,the feature of the object is designed manually,which has the characteristics of long engineering time and poor robustness.With the development of deep learning,high dimensional features of images can be extracted efficiently through convolutional neural network.Nowadays,among the methods of image detection and recognition,deep learning has become the most popular and practical.Based on the analysis of Faster R-CNN algorithm and YOLOV3 algorithm and the theoretical technology of deep learning,this paper proposes corresponding improvement strategies to realize vehicle image recognition and detection.The main work of this paper is as follows:(1)In this paper,by improving the Faster R-CNN algorithm to learn the building database image,Through the K-Means++ clustering method,the positions of known target boxes in the training set were clustered,and K different clustering centers were set through experiments.analyze the best clustering results,implement the algorithm training on the training set image through the Tensorflow deep learning framework,replace the set 9candidate Windows with K clustering centers with good effect,and improve the detection speed through experimental comparison.(2)In this paper we also propose an improved YOLOV3 method based on multi-scale fusion to improve the detection accuracy of small target objects.Firstly,the improved K-Means clustering algorithm is used to sample the K-Means clustering method and the kernel function to cluster the dimension of the number of width-to-height ratio of the target candidate box,and multi-scale fusion is carried out for the shallow feature information of the small target object.YOLOV3 is improved by adding two kinds of multi-scale fusion.Then,the improved YOLOV3 algorithm is used to carry out experimental comparison with YOLOV3 on the KITTI data set.The simulation results show that the improved network can effectively improve the detection effect of small target objects,The recall rate and accuracy of small objects have been improved.