Research On Object Detection Based On Deep Learning

The rapid development of China's economy has led to a continuous increase in the number of motor vehicles,which has also triggered a series of problems.Therefore,the Intelligent Transport System(ITS)attracts more and more attention,and such systems can increase the safety of vehicle driving.Vehicle object detection is one of the core functions of the ITS system.The traditional methods of video surveillance image processing have achieved great development,but the complex background,strong light,rain and snow and other bad weather and video surveillance image overlay in the vehicle,When these problems exist,the detection effect of the traditional methods is still not good.This article summarizes the traditional vehicle detection methods.For its shortcomings,deep learning based detection algorithms are applied to vehicle detection.Focusing on the solution to traditional methods does not address the problem of locating vehicle targets and classifying vehicle targets.Mainly completed the following work:1)This article produced a self-built vehicle data set for experiments.The experimental picture was obtained by the web crawler and the camera photographing.A variety of vehicle images with different appearances,models,and angles were collected from different traffic scenes.Using Image Net annotation to form a standard vehicle data set.2)Under the dark light environment,the vehicle object and background in the image are similar,and it is difficult to distinguish between.Therefore,this paper proposes an algorithm for judging vehicle images under dark light conditions,which makes it possible to distinguish between normal light images and dark light vehicle images,and to enhance the processing of dark light vehicle images.Making better separation of the treated vehicle object and background,and finally use the convolution network to locate and classify vehicle targets..3)In the process of vehicle object positioning,the positioning of the bounding box of the vehicle object in the convolutional network is inaccurate.This paper improves Loc Net convolutional network vehicle location algorithm and obtains more search areas through different expansion factors.Makes it possible to achieve better vehicle object positioning in subsequent convolution processing.4)In order to better match vehicle detection in real-life scenarios,this paper improves the SSD convolutional neural network.It is possible to implement detection and vehicle model recognition for multiple scales of vehicle objects in multiple categories in the scene.Applied to the detection of electric vehicles for the first time,and the vehicle detection method also has good test results in fog,cloudy,rain and snow conditions.