Anchor Generation Algorithm And Small Obstacle Detection Network Research

In the field of autonomous driving,the detection of small obstacles on the road can help cars avoid obstacles and ensure driving safety.With the promotion and popularization of new energy vehicles with assisted driving functions,the detection and positioning of small obstacles is becoming more and more meaningful.However,small objects occupy few pixels in the image and the edge information is blurred,which make the research on small object detection full of challenges.In this thesis,a joint dataset of small targets and obstacles on the road is created.The dataset is rich in driving scenes,diverse in target types,and small targets account for more than 90%.The performance of the target detection algorithm on the small target joint dataset is used to measure the performance of the algorithm for small target detection.In this thesis,the singlestage target detection algorithm is used as the main frame of the target detection network,and the experimental research on the detection of small-sized targets is carried out around the distribution characteristics of the small obstacle samples and the neural network structure.The specific work has the following two directions:(1)In terms of anchor box generation,during the implementation of the single-stage target detection algorithm,it is necessary to generate an anchor box,that is,a priori box,according to the data set,which is generally generated by the K-means clustering algorithm.By studying the properties of the intersection and ratio,this thesis proposes concepts such as point symmetry,line symmetry and point-to-line intersection distance under the intersection distance.On this basis,in the process of generating anchor boxes by the K-means algorithm,the random generation of initial values can not ensure the quality of the anchor boxes,the clustering process is sensitive to the data aggregation area,which may easily lead to dense distribution of anchor boxes,and the use of mean to generate cluster centers is not an optimal solution.In order to solve the problems,a guided K-means clustering algorithm is proposed.Initializing the anchor frame according to the guideline reduces the randomness of the anchor frame generation,using the guideline to constrain the range of the anchor frame during the clustering process alleviates the problem that the anchor frame is too dense due to the sample distribution,according to the average intersection of the samples and the nature of the convex function.The optimal solution is obtained by solving the cluster center using the hillclimbing algorithm.Finally,high-quality anchor boxes can be stably generated according to the dataset,which improves the performance of the target detection network to a certain extent.(2)An improved feature pyramid network structure FPNS is proposed for small obstacle targets.By studying the optimal number of anchor boxes for small obstacle data,it is found that the network performance of two layers of detection layers is better than that of three layers of detection layers.It is also found that in the case of a single anchor frame in the single-layer detection layer,the target algorithm has a very poor effect on the detection of small samples whose intersection ratio with the anchor frame is less than 0.25.Therefore,an input four-layer feature map and output two-layer feature map to be detected are proposed.Because of the fusion of high-resolution shallow features,the detection problem of extremely small samples is solved,and the detection performance of the target detection algorithm is improved.In the end,a recall rate of 97% and an accuracy rate of 90% are achieved.