Research On Target Detection Algorithm Combined With Scene Semantic Information

In recent years,object detection,as the core technology of visual information acquisition for driving aid system,has important practical significance for automatic detection of road obstacles.In the real driving scene,because the car is usually driving in the road area,the obstacles in the sidewalk area have no impact on the driving safety,so it is very important to obtain the area information of the obstacles.In this paper,based on the research of obstacle detection,the scene semantic information is obtained by combining with semantic segmentation,and the research of object detection combined with scene semantic information is carried out.(1)Aiming at the problem of low detection accuracy of YOLO v4,we design an improved algorithm.Kmeans++is used to generate anchor boxs suitable for the data set to enhance the scale adaptability.CIoU is used as the loss function of coordinate prediction to improve the accuracy.Finally,the improved YOLO v4 detection algorithm is obtained.Experimental results show that the detection accuracy is improved by 2.03%compared with YOLO v4.(2)Aiming at the problem that the semantic segmentation model of Deeplab v3+algorithm is not accurate enough,this paper improves the network structure of Deeplab v3+algorithm.The context path module is introduced in parallel with the original ASPP module to increase the receptive field of the network,so as to improve the detection accuracy of large targets.Experimental results show that,compared with the classical segmentation algorithms SegNet and Deeplab v3+,the average intersection ratio of the improved Deeplab v3+algorithm increases by 2.02%and 0.14%respectively,the intersection ratio of roads increases by 0.62%and 1.76%respectively,and the intersection ratio of sidewalks increases by 1.80%and 7.77%respectively.(3)In order to obtain the region information of obstacles,this paper proposes a target detection algorithm combined with scene semantic information.Firstly,the obstacles in the streetscape are detected to obtain the target classification and prediction information of the detection frame;secondly,the streetscape is semantically segmented to obtain the category of each pixel to obtain the pixel information of the road and sidewalk area;finally,by synthesizing the above information,the pixel information in each obstacle detection frame is counted to obtain the area information of the obstacles.According to the statistics of the proportion of road pixels in the range of 70%,40%and 10%below the detection frame,the average precision rate of road obstacles is 95.15%,96.14%and 92.68%,and the average false alarm rate is 4.80%、4.18%and 7.72%.The average processing time of single image is 0.67s.The experimental results show that the semantic information of a single object can be obtained effectively according to the pixel statistics of the lower 40%of the detection frame.To sum up,the improved Yolo V4 target detection method can detect obstacles more accurately;the improved Deeplab v3+semantic segmentation method can improve the detection accuracy of road and sidewalk area pixels;based on this,a target detection algorithm combined with scene semantic information is proposed,which can effectively obtain the area information of obstacles,so as to distinguish the obstacles that have an impact on driving safety.In this paper,combined with the existing research content in the field of assisted driving,a method to obtain road obstacle information is proposed,which is of great significance to improve driving safety.