Research On Intelligent Vehicle Road Detection Technology Based On Machine Vision

With the development of social economy and the progress of automobile industry in our country,the demand and quantity of automobile have repeatedly reached new highs.The development of intelligent automobile technology has become the consensus of automobile industry and researchers.Road detection technology is an important part of intelligent vehicle environment perception.Based on machine vision,this paper studies road detection technology and takes accuracy and real-time performance of road detection as key indicators of the algorithm.Specific contents include:(1)Road image acquisition and preprocessing.Based on the intelligent vehicle platform,the road image is collected,and the region of interest division method based on the gray mean of image lines is proposed,and the image data is preprocessed by combining the calibration of vision sensor and the inverse perspective transform.(2)Structured road detection.In this paper,structured road detection is transformed into lane detection.Lane color,gradient and edge feature were used as targets to filter lane line pixels.The distribution of lane line pixels in the graph was calculated,the specific coordinate position of lane line pixels was obtained based on the sliding window method,and the lane lines were fitted with the least square method to detect the single frame road image.To solve the problem of lane discontinuity,an interframe tracking algorithm based on Kalman filter was proposed to improve the accuracy of continuous frame detection.Experiments show that the proposed algorithm takes 68 ms to detect a single frame,has good accuracy and real-time performance,and has certain robustness in the face of complex road conditions and interference conditions.(3)Unstructured road semantic segmentation.Intelligent vehicle road detection algorithm needs to balance accuracy and real-time performance.Based on this requirement,DeepLab V3 network structure is improved,lightweight network MobileNet＿V2 is used to replace part of ResNet structure for feature extraction for lightweight design.In view of the accuracy decline caused by improved network structure,The effect of activation function on neural network performance is analyzed and Swish nonlinear activation function is introduced to compensate the accuracy.The experiment shows that the semantic segmentation network established in this paper achieves 71.57% overlap rate in Cityscapes dataset and 84% single overlap rate in road areas.The single frame detection time is 54 ms,which is 96% less than DeepLab V3+network parameter and 77.1% less than Deeplab V3+ network parameter.Achieve the performance of DeepLab V3 network.(4)Division of passable areas.The maximum inter-class variance method is used to obtain the road region in the semantic segmentation image and the region growing method is used to eliminate the false detection region and disconnected region in the road region.Finally,a regional optimization algorithm was established to eliminate the spikes and edge error detection pixels in the road area,and the passable area was divided by taking the width of the video vehicle as the threshold.The experiment shows that the single frame time of the road area optimization algorithm is 9ms,and the intersection rate of road areas increases by 7.33%after optimization,reaching the level of road area segmentation of DeepLab V3+.Experiments show that the road detection algorithm established in this paper has certain accuracy and real-time,and has certain application value,which can provide theoretical support and method reference for the subsequent research in the field of intelligent vehicle environment perception.