Research On Real-time Detection System Of Vehicle Line Pressing Based On CNN

The illegal lane changing behavior of vehicles has become an important factor affecting traffic safety.At present,the vehicle line pressing detection device is still based on traditional digital image processing methods,combining with geomagnetic coils,acoustic velocimeter and other hardware sensors to identify vehicle behavior.However,the traditional digital image processing methods are greatly affected by environmental lighting,climate conditions and other factors,and the algorithm robustness is poor.The laying of hardware sensors will damage the road surface The flexibility is also very poor,which reduces the work efficiency of China’s traffic management departments and increases more review costs.In recent years,the target detection method based on deep network has developed vigorously.Applying it to the vehicle behavior detection at the road intersections in China can effectively reduce the project cost and further improve the intelligent progress of the traffic management system in China.The main research work of this paper includes:Aiming at the low detection accuracy and poor robustness of traditional digital image processing methods,this paper uses the target detection algorithm based on depth convolution neural network to detect the location information and category information of vehicles.Based on YOLOv5 s general target detection algorithm,combined with specific application scenarios and vehicle category information,a vehicle detection algorithm dedicated to road traffic intersections is designed.The network model is optimized by cutting redundant small target output features in actual application scenarios.In view of the lack of vehicle feature information in the public data set,the data set is expanded.After training,the network model is evaluated with the indexes of accuracy rate,recall rate,average accuracy and real-time detection speed,which shows that the network model has a good detection effect on vehicles.For the detection of lane lines,considering the practicability of the algorithm and the consumption of hardware resources,the traditional digital image processing method is used to extract the lane solid line contour.First,filter the scene of the road intersection to remove the impact of related noise,and gray and binary processing the image to merge the pixel channels and highlight the lane line information.Then the edge feature information of the lane solid line is extracted,and the maximum area search is selected to extract the lane solid line information after comparing the Hough transform and contour search methods.The vehicle position information obtained from the regression of CNN detection network is calculated to simulate the four coordinate points of the wheel.The minimum distance between the coordinate points and the solid line contour of the lane is calculated to determine whether the vehicle has line pressing behavior.The ZYNQ heterogeneous platform used in the experiment completed the hardware deployment of the CNN based real-time detection algorithm for vehicle crimping,and combined the control advantages of the microprocessor with the computing advantages of programmable logic to complete the entire system.Aiming at the problem of large network model volume on the computer side,the network model volume is compressed by pruning and quantifying methods.The experimental data shows that the network processing speed on the AXU2 CG side is 16.62 frames per second,and the average accuracy is 79.6%,which meets the real-time detection speed and accuracy requirements of the application scenarios at the traffic ports.Figure [77] Table [6] Reference [64]...