Research On Vision Assistant System For Closed Road Driving In Haze

With the acceleration of industrialization and modernization,the expansion of traffic infrastructure,and the increasing number of cars,traffic safety accidents have become a concerning issue.Therefore,it is crucial to study relevant driving assistance systems that can detect and warn of potential dangers for daily driving safety.This study focuses on driving on closed roads during haze days and proposes a video frame dehazing algorithm based on the dark channel priori and a road condition driving detection algorithm based on YOLOv7.These algorithms are implemented on a CPU-GPU heterogeneous platform to develop a visual aid system for driving on closed roads during haze days.The contributions of this thesis are as follows:(1).The study proposes a multi-transmittance fitting algorithm to address the halo effect and color distortion of the dark channel prior algorithm.A new method for calculating global atmospheric light value is also proposed.The algorithm’s speed is accelerated by using the nonlinear programming solution model and skewness theory on a CPU-GPU heterogeneous platform.(2).To address the dark channel prior algorithm’s failure to process the sky region,the study calculates the confidence of the video frame using the brightness V and saturation S spatial information of the two-dimensional Gaussian function and HSV color model.An image fusion and enhancement strategy is also designed to enhance the feature information for subsequent target detection.The algorithm’s speed is improved on a CPU-GPU heterogeneous platform.(3).The study proposes a traffic detection algorithm based on YOLOv7.The convolution block attention module is added to the backbone network to improve its ability to extract target features.The traffic detection rate is improved under haze road conditions on closed roads by improving several MPConv layer modules in the network structure.Tensor RT is used on a CPUGPU heterogeneous platform to realize detection acceleration.The study conducts experiments and analysis on dehazing effect,dehazing time,model training,and target detection.The results demonstrate that the proposed algorithm has good dehazing effect,rich scene feature information,and real-time dehazing performance.The realtime reasoning and detection effect of driving targets are good.The algorithm also has significant advantages in visible edge increase rate,information entropy,average gradient,dehazing time,m AP@0.5 Value,target detection rate,and reasoning detection time compared to recent relevant algorithms.