Research And Implementation Of Driving Early Warning Algorithm Based On Deep Learning Image Processing

The driving early warning system based on the deep learning pre-processing of the monocular camera needs to prompt the driver of dangerous information based on environmental perception.Pavement basic environmental information should include vehicle information and lane information.The system uses a low-cost monocular camera to obtain environmental information,and then obtains the information of the vehicle in front and the lane information of the vehicle through the deep learning pre-processing module,and finally combines the early warning and post-processing module for analysis to obtain warning information of dangerous road conditions.In order to avoid the problems of high time-consuming,high false detection rate and low accuracy caused by traditional computer vision,a driving warning system scheme based on deep learning pre-processing of monocular camera is proposed.The use of low-cost cameras and the ability to quickly infer with deep learning provide timely and accurate warning information to the driver to avoid traffic accidents in a timely and effective manner.First use the CoCo2017 data set to pre-train the network.Then created a CQ＿vehicle＿dataset database specifically for vehicle detection,and trained the pre-trained weights again.In order to make the network more suitable for the detection task of vehicle targets,Kmeans ++ algorithm was used to perform cluster analysis on the aspect ratio of 66389 different vehicle targets.Based on the original Mask R-CNN network architecture,a set of Anchor ratios are added to improve the detection accuracy of vehicle targets.For the detection of lanes,an end-to-end training method is used to directly obtain the second-order fitting coefficients of the left and right lanes of the lane.It is verified that the process of fitting the lane line using the least square method is differentiable and can also be propagated back in the network.On the basis of the existing labeling information in the Tusimple Dataset dataset,the labeling information of the lane line fitting coefficients is added,and then end-to-end training is performed.The mean square error loss and geometric loss function were compared experimentally,and geometric loss was selected as the loss function.The network output is pre-processed and then used as the input to the post-warning postprocessing module.The early warning post-processing module takes the vehicle and lane information output from the two networks as input.Two modules of collision warning module and front lane change warning are designed.The collision warning module first screens the vehicle target,and then designs a first-person distance measurement method based on the internal parameters and installation parameters of the monocular camera,which achieves accurate distance measurement of the vehicle target.According to the relationship between the vehicle and lane angle change rate,the front lane change warning module implements a first-person perspective lane change behavior discrimination system,and achieves good results.Combining the two early warning sub-modules,the driver can be promptly and accurately informed of the risks.