Visual Recognition Algorithm For Mobile Scenes And Its Application In Urban Management

Illegal sidewalk parking has traffic safety hazards,which is a difficult point in urban management.The traditional surveillance camera is not suitable for mobile scenes,while manual inspection has the disadvantage of low efficiency.Aiming at the above problems,this thesis proposes a visual recognition algorithm for illegal parking car based on the theory of deep learning,which improves the effect of instance segmentation network on the image captured by mobile devices.The main contents of this thesis are as follows:(1)Establish urban sidewalk parking scenarios dataset.Flip,rotate and zoom the images in order to increase the size of dataset and enhance the generalization of model.Considering that it will take too much time in manually solving pixel-level labeling tasks,design fast labeling process based on Polygon-RNN model.(2)Design recognition method for parking area based on instance segmentation network model Mask R-CNN.According to dense connection,optimize the feature transmission mode in basic network structure and strengthen positioning feature information of large objects.Adjust the number of anchor and candidate proposal boxes in the proposed area in order to reduce the complexity of the model.(3)Optimize Mask R-CNN instance segmentation process and design illegal parking determination method.Apply Soft-NMS to solve overlap problem,and use fusion method for connected region based on Mask-Io U in order to improve the instance segmentation results of parking area.After the optimization of post-processing,design illegal parking determination algorithm by calculating the overlap rate between the car and parking area and analyzing their relative positions.The algorithm effectively distinguishes the legal parking cars with the illegal ones and has achieved a recall rate and a precision rate of 90.5% and 80.8% for the car on the sidewalk respectively.(4)Summarize the results and put forward the problems of algorithm that need to be solved in the future.The algorithm proposed in this thesis significantly improves the recall rate and precision rate of illegal sidewalk parking,effectively solves the problem of traditional surveillance camera detection and manual inspection.It helps to realize the intelligent detection of illegal sidewalk parking and improves the intelligent level of urban management.