The Free Space Detection Technology Research For Urban Road Automatic Driving

With the rapid economic development in our country,the number of motor vehicles has been increasing year by year.While the car brings tremendous convenience to our life,the traffic safety problems that arise are more and more intuitively presented to us.There is a survey shows that the vast majority of traffic accidents with the driver's driving status,driving habits,the emergence of automatic driving technology has become a revolutionary solution to the above problems.Environmental perception is the research focus of autopilot,and as an important means of environmental perception,the image has become the hotspot of research because of its abundant information and real-time.Semantic segmentation techniques can provide critical information such as the contour of an obstacle,the boundaries of drivable areas.In recent years,with the rise of the deep learning wave,the performance of deep neural networks in image classification and target detection has been outstanding,showing great advantages compared with traditional machine learning.In the field of automatic driving,deep learning has become an important research direction of environmental perception and image segmentation.This paper mainly studies the free space detection technology for urban road automatic driving,the main contents are as follows:1.Based on the classification of residual networks with excellent performance,this paper studies the fully convolutional technique that using the network model for image semantic segmentation.It turns the one-dimensional label output of the classification network into a two-dimensional output of the feature map.In order to make the size of the segmentation result equal to the original image,we take a deconvolutional layer.In terms of "what" and "where" in image semantic information,the method of feature fusion is combined with the superficial information of the network to optimize the segmentation result;2.In order to solve the problem of misclassification of the image boundary region in the fully convolutional network,we add the key feature extracted by human in the training phase of the semantic segmentation network,adopt the method of parallel structure and multi-layer fusion.Compared with a single network structure,this parallel structure has a better performane.The conditional random field algorithm is used to optimize the segmentation boundary at the output of the network.Experiments show that our algorithm makes the segmentation more clear and precise;3.We validate the segmentation of the two models on a urban traffic scene dataset that is labeled pixelwise.In order to highlight the segmentation of the free space,we separate the different semantic area in the segmentation map into research experiments.We report three metrics from common semantic segmentation and scene parsing evaluations that are variations on pixel accuracy,mean pixel accuracy and region intersection over union.The results show that our proposed model segmentation is quite effective.