Research Of Lane Detection Algorithm Based On Machine Vision

In recent years,many machine vision task's problems can be solved by deep learning method.The prediction results have better robustness and generalization.An image pre-processing method is designed for lane line detection aim at the problem of noise interference and image perspective relationship in the process of lane image.Firstly,the region of interest of the road image is extracted according to the vanishing point' position to remove the image area unrelated to the road;secondly,images can be filtered out the noise caused by camera imaging through the filter;finally,the lane perspective images are transformed into the top view of lane images by inverse perspective transformation to eliminate the impact of perspective relationship for improving the detection accuracy and speed.In order to improve the performance of lane line detection,a semantic segmentation algorithm with improve Oct Conv for lane is proposed.It combines the FCN network with image pre-processing methods to fulfill a good detection of lane line.The algorithm is used to divide the whole image into high-frequency components and low-frequency components for lane line feature extraction according to the distinct features of lane image and the redundancy of non-lane image.Then combining with the dilated convolution,the loss of image information in Oct Conv is improved to achieve the end-to-end detection of lane image,which improves the detection accuracy without sacrificing the real-time performance.In order to obtain the function of lane,a method is proposed which uses the least squares to fit the lane and get the lane function according to the geometric characteristics of lane line.It has a better detection effect on the scene images of vehicle occlusion,curve,lane change and shadow,then the fit lane line has the ability to ignore vehicle occlusion.Compared with other algorithms,the overall performance of lane line detection is better than other algorithms in the experimental data.The experimental results show that the presented method can effectively improve lane detection performance and has better scalability.