A Study On Vehicle Detection And Road Segmentation Technologies In Fixed-point Monitoring Viewport

The acceleration of urbanization has brought about problems of illegal land occupation and illegal construction,seriously affecting national food security and social sustainable development.Compared with manual inspection,it is more efficient to detect illegal buildings through detecting building changes between images.However,the changes of roads and cars in the image make the building change detection more difficult.To overcome the problem,the paper studies vehicle detection and road segmentation techniques in fixed-point monitoring viewport based on deep learning to remove the interference,so as to support early detection and processing of illegal construction.The main research contents are as follows:(1)A DSP-YOLO object detection algorithm based on dilated spatial pyramid is proposed.Aiming at the problem of poor multi-scale detection performance of YOLOv3,a novel dilated spatial pyramid(DSP)module is designed to integrate multi-scale information by re-sampling the feature maps with parallel dilated convolution branch.In view of the problem of that vehicles under the fixed-point monitoring perspective are generally small,feature maps with larger resolutions are fused.Experiments show that DSP module improves the multi-scale object detection performance and achieves 94.5% m AP on the vehicle dataset.(2)A DC-Deep Lab semantic segmentation algorithm based on dense connection and channel attention is proposed.According to the problem that deeplabv3+ performs poor on edge segmentation,three multi-scale context information encoding modules are designed to provide more effective context information.After that,a channel attention mechanism is introduced in the decoder module to pay more attention to object information and suppress noise,making the segmentation result more refined.Experiments show that the proposed DC-Deep Lab algorithm effectively improves the segmentation performance,and achieves89.16% MIo U on road dataset.(3)An improved DC-Deep Lab road segmentation algorithm based on progressive feature fusion is proposed.Considering the problem of insufficient use of low-level features in the decoder module of DC-Deep Lab,a progressive feature fusion module with dual attention modual is designed to fully intergrate low-level features.The improved DC-deep Lab achieved 90.85% MIo U on the task of road segmentation.