Research On Key Technologies Of Obstacle Detection In Complex Urban Roads Based On Edge

Auto-assisted driving technology has always been a hot topic in the field of machine vision.Obstacle detection is the key to auto-assisted driving technology and provides an important guarantee for safe driving of vehicles.Due to the complexity of the surrounding environment in urban roads,especially in downtown areas,the detection of obstacles in roads in complex urban areas is a difficult point in the detection of obstacles,which poses a great challenge to the detection of obstacles.This paper conducts in-depth research on this issue and proposes a method for detecting obstacles in complex urban roads based on edges.The specific research content includes the following aspects:Firstly,based on K-Means,an unsupervised feature learning road region detection method is proposed for complex urban roads.Firstly,shadowed road maps are de-shaded,roads prior knowledge and histogram normalization methods are used to extract road confidence regions,and K-Means clustering algorithm is used to characterize pre-processed road confidence maps.Then,the learned features are scored based on the position information of the road image,and the score is compared with the front feature map.The feature map is segmented by using multi-threshold method to obtain the road region.Then,an improved Edge-Boxes obstacle candidate region extraction method is proposed.The algorithm obtains the edge of the image based on the gradient information,clusters the edge pixels according to the gradient value and direction into individual edge group sets,then uses the edge group set to score the candidate bounding box and according to the target outline in the candidate area border.Number of obstacle candidate area extraction.Finally,pedestrians and vehicles are detected using a convolutional neural network.The normalized target candidate region is input into the convolutional neural network,and the feature vector corresponding to each candidate region frame is output.Then the SVM classification algorithm is used to classify the extracted features to separate the target to be detected.In order to verify the performance of the proposed algorithm,the algorithm was tested on three road data sets.The experimental results show that the proposed road detection algorithm has high robustness,with the accuracy and recall rate of 84% and 90% respectively.The improved target candidate region extraction algorithm avoids the time spent training the model,and the test results on the VOC2007 data set show that the recall rate is still around 80% at IoU of 0.6;At the same time,this paper successfully combines the improved target candidate region algorithm with convolutional neural network to effectively detect obstacles in complex road scenes on the KITTI data set.