Research On Road Obstacle Detection Based On Deep Learning

With the vigorous development of road construction in China,the traffic between cities and cities is becoming more and more developed,and the road safety detection and maintenance work is increasing.However,the traditional manual inspection work is inefficient,but at home and abroad.The research on road disasters is mainly reflected in the analysis of road structure and the prediction of hazards.Therefore,according to the deep learning theory,this paper designs a method based on deep learning to automatically monitor road obstacles,which can automatically identify obstacles on the road and automatically classify and mark their specific positions,realizing automatic marking and processing of a large number of road pictures.Can effectively improve the efficiency of road safety monitoring.Aiming at the problem of the collection and production of road obstacle image dataset,this paper designs a set of automatic dataset generation methods.First,use the web crawler technology to crawl the images of interest on the network,and automatically perform image scaling,enhancement,noise reduction,data augmentation,etc.Finally,manually filter the images and mark the object.Using this method,we collected 3000 positive samples and 300 negative samples as the data set of this experiment.Then the paper introduces the Mask R-CNN object recognition network,and proposes the improved Mask R-CNN network based on the fusion BN layer and convolution layer,changes the mask prediction branch,etc.,and uses the data set collection method designed in this paper.The collected data set details the process of the improved Mask R-CNN network structure and training parameters,and the final network model is obtained.The experiment found that the improved Mask R-CNN network structure designed in this paper increases the recognition speed by about 15% without reducing the accuracy.At the same time,the network has a accuracy of about 90%,and can automatically mark the location and type information of road obstacles in the picture.