Research On Vehicle Detection Algorithms Based On Deep Learning

With the development of the science and technology,human beings have gradually entered a world of networking and intelligence.In the era of Artificial Intelligence,which takes advantage of deep learning,automatic driving has achieved unprecedented attention.In the "New Generation of Artificial Intelligence Development Plan" issued by the State Council,the research of automatic driving technology has risen to the national strategy.Automatic driving combines four basic elements: big data,algorithm and decision-making,computing power and sensor data acquisition.It has important applications in military defense and daily life fields.Object detection algorithm is one of the basic supporting technologies of automatic driving.Automatic driving requires precise positioning of surrounding objects to guide vehicle routing planning,avoid obstacles and ensure driving safety.Object detection usually consists of vehicles,pedestrians,traffic signs and other objects.According to the different information obtained,it can be divided into 2D object detection and 3D object detection.Traditional object detection algorithms adopt sliding window strategy,and the detection efficiency is very low;moreover,feature extraction relies on manual adjustment,and the ability of feature expression is very poor,so it is unable to extract effective semantic information.With the development of deep learning,data adaptive convolution kernels can make full use of texture,color and shape information in images,which greatly improves the efficiency and accuracy of object detection.Therefore,this paper uses deep learning method to carry out research on vehicle detection.The main work is summarized as follows:(1)The current development status of object detection at home and abroad is summarized,and the technical bottlenecks of object detection algorithm at the current stage is analysed.A detailed introduction of KITTI dataset and the evaluation indicators such as Intersection over Union and the accuracy rate is provided.(2)To solve the problem of low recall and inaccurate location of small objects in Faster R-CNN algorithm,this paper proposes an improved 2D car detection algorithm from two aspects: multi-scale feature and detector cascade.In order to make full use of the high resolution and precise location information in shallow features and the high semantic information in deep features,a top-down structure with horizontal connection is constructed.At the same time,the Anchor ratio is extended to improve the robustness of the model to vehicle shape change.Finally,the detector is cascaded to improve the detection performance of the model by increasing the detector threshold.Experiments on KITTI vehicle detection data set verify the effectiveness of this method.(3)Aiming at the problem that the 3D coordinate information of objects can not be obtained directly from monocular images,this paper proposes an end-to-end 3D vehicle detection algorithm based on the geometric constraints provided by perspective projection.In geometric constraints,the key point constraints are adopted in this algorithm,because the camera’s internal parameters and the object’s 3D bounding box are given in the training data set,so the eight vertex coordinates of the 3D bounding box can be directly obtained in the image through the mapping matrix.The mapped key points are closely related to the local features of the object,and can be located by the key point detection algorithm.The above information can be estimated simultaneously by the convolutional neural network designed in this paper.This algorithm is compared with other similar algorithms on KITTI vehicle detection data set,and has better detection effect.Finally,our work is summarized and we look forward to the future research work.