Research On Object Detection Algorithm Based On Single Shot Multibox Detector

In recent years,with the development of computer software and hardware technology,deep learning and convolutional neural network technologies have set off a wave in the field of computer vision.The emergence of convolutional neural networks has solved the problem of object feature design in images.Convolutional neural networks have been applied in more and more studies in the field of computer vision.Object detection,as a basic task in the field of computer vision,has been the subject of popular research.Domestic and foreign scholars have proposed a variety of object detection algorithms based on convolutional neural networks and and their improved versions,such as R-CNN,Fast R-CNN,Faster R-CNN,RFCN,Mask RCNN,Yolo series,SSD and other object detection algorithms.Single Shot Multi Box Detector(SSD)is a one-step detector based on deep convolutional neural network.It has the advantages of fast detection speed and accuracy comparable to that of two-step detector.However,the SSD object detection algorithm still has problems such as poor detection of small objects,single feature extraction,and difficulty in real-time detection of objects on ordinary machines.These problems limit the performance of SSD object detection algorithms.Therefore,this thesis proposes an improvement scheme for the shortcomings of the SSD object detection algorithm.The main work of this thesis is as follows:1.Aiming at the single feature extraction method of SSD object detection algorithm,this thesis proposes a SSD object detection algorithm based on multi-scale convolution structure(MSSD)to further improve the detection accuracy of the algorithm.The convolution kernels of different sizes are arranged side by side to form a multi-scale convolution structure and added to the detection layer of the SSD target detection algorithm to enrich the diversity of feature extraction.Simulation experiments were carried out on the PASCAL VOC public dataset.The average accuracy of the various objects(Average Precision,AP)and dataset average accuracy(mean Average Precision)obtained by the original algorithm and the proposed improved algorithm were analyzed and compared.simulation results show that the proposed multi-scale convolution structure improves the performance of target detection accuracy,m AP increased from 75.3%to 76.4%,increased by 1.1%,but the rate of detection frames per second(fps)decreased from 16.5 frames per second to 13.4 frames per second.2.Aiming at the problem of real-time detection of target detection for MSSD algorithm,an improved scheme of prior boxes extraction based on pooling feature is proposed without affecting the accuracy of algorithm detection.The convolution feature map is effectively compressed into a pooled feature map and detected for the pooled feature map.As the size of the detection feature map is reduced,the number of priori boxes is also reduced.In the simulation,the proposed priori boxes extraction improvement scheme is combined with the MSSD algorithm.The simulation results show that the improved algorithm improves the m AP from 75.3% of the original SSD object detection algorithm to 75.9%,and the fps is 16.5 frames per second has been increased to 19.1 frames per second,and the detection accuracy and speed have been improved at the same time.3.Aiming at the problems of different illumination conditions in traffic scenes,long-distance vehicles in different backgrounds,and real-time detection of small traffic lights,a practical design scheme for realtime detection of objectss in a real-world scenario is proposed.A dataset that requires human marking is used in the actual traffic scene,and the dataset contains information such as vehicles and traffic lights.Based on the lightweight convolutional neural network Mobile Net and the idea of SSD target detection algorithm,a real-time detection of lightweight network structure—Micro Convolutional Neural Networks(MCNN)is proposed and designed.Finally,the simulation experiment is carried out in the traffic dataset,which realizes the detection of small vehicles and traffic lights under different illumination conditions and long distances.The detection speed can reach 27.1 frames per second,which verifies that the design can perform real-time testing.