Research On Road Detection Based On Machine Vision

Driverless are very important in reducing traffic accidents,alleviating traffic pressures,changing travel modes and promoting automotive industry development.Road detection technology based on machine vision is one of the most important technologies for driverless vehicle.In this paper,road detection based on machine vision was researched.Road detection is divided into two parts: structured road detection and unstructured road detection.Structured roads detected by lane feature which extracted through the image.Unstructured road detection uses area segmentation based on image semantic segmentation.Structured roads usually have clear and fixed-style lane.Its use lane detection instead of road area detection.Firstly,according to the gradient,direction and distribution characteristics of the lane to get the edge of lane.Secondly,possible lane areas are filtered through HLS color modes L and S channels.Then,the lane edge features and regional features are fused.Combining the relationship between the width of lane and vehicle position,the real position of lane are obtained.Finally,the lane fitted by quadratic polynomial.The experiment shows that the algorithm has good realtime performance,the detection time of single frame is 73 ms.Under good weather conditions,it has good anti-interference ability to words and direction signs in the lane.Unstructured road area extracted by image semantic segmentation model.An image semantic segmentation network based on deep learning model is designed for road detection.The model architecture is encoder-decoder.The model encoder uses a lightweight feature extraction network MobileNet_V2,and the decoder uses a low-level,high-level feature fusion network based on multi-level upsampling.Mean intersection over union is used as performance evaluation criterion.The design network achieved mean intersection over union of 75.67% and 71.63% at PASCAL VOC2012 and Cityscapes datasets,respectively.The lightweight design network model reaches the level of the first generation DeepLab image semantic segmentation network.After the road segmentation is completed,the road area optimization algorithm is used to optimize the road area.Firstly,the region growing method is used to determine the current driving area of the vehicle.Then,the inverse perspective transformation is applied to restore the road spatial information.Finally,the row and column edge burr optimization algorithm is used to optimize the edge burr and improve the overlap rate of road area.The experimental results show that the road intersection over union is 94.41%,which reaches DeepLabV3+ level.Meanwhile,the single frame use time less than 52 ms,which is 80% lower than DeepLabV3+.The lane detection and road area segmentation algorithm designed in this paper have good application value and can provide theoretical support and method support for the development of driverless road detection technology.