Road Detection Based On Density Peaks Clustering And Fully Convolutional Networks

For autopilot systems,accurate and real-time road detection algorithm is the core technology.The basic principle of the road detection algorithm is to capture the surroundings of the vehicle through the camera fixed on the vehicle,and then use the sensor to process the environmental information so as to detect the road area that can be driven in front of the vehicle and provide assistance for vehicle path planning and road sign detection.However,due to the complexity of the environment,the algorithm results are easily affected by various noises,such as uneven lighting,pedestrians,vehicles,and shadows.Therefore,when the environment is more complex,how to improve the accuracy and robustness of the algorithm has become people's research hotspots.This thesis mainly studies the detection of road areas under complex scenes,especially when there are shadows,uneven illumination,pedestrians,vehicles and other interference factors in the environment.The main research of this thesis is as follows:(1)In view of the existence of shadows in the image,this thesis proposes an image preprocessing algorithm based on color features and entropy theory.Firstly,the color feature pairs of the image are extracted.Secondly,these color feature pairs are projected from different perspectives to convert them from two-dimensional space to one-dimensional space and the amount of computation can be reduced.Finally,the image reconstruction is completed based on the entropy theory.Experiments show that the proposed algorithm can not only remove most of the shadows in the image,but also clearly show the outline of the road and sky.(2)This thesis proposes a dividing line detection algorithm based on image connected component labels,aiming to remove the impact of sky on road detection.The algorithm makes full use of the features of the preprocessed image,which can clearly show the outline of the road and sky.Firstly,the preprocessed image is transformed into a binary image and the road scene is simplified.Secondly,the morphological operations are used to remove the noise in the image.Finally,the maximum connected region of the image is marked to detect the dividing line between sky and road.After experiments on three different databases: After-Rain,Sunny-Shadows and Baidu pictures,qualitative and quantitative results prove that the proposed algorithm has higher accuracy and it can detect the dividing line in the image more accurately,thus removing sky's interference with the road detection algorithm.(3)This thesis proposes a road detection algorithm based on density peaks clustering.Firstly,the original image and the image marked with connected region are fused,and the gray value features and luminance value features of the image are extracted.Secondly,the image below the dividing line is cut into multiple sub-images at certain intervals,and each sub-graph is clustered using the density peaks clustering algorithm.Finally,the road classes in the subgraphs are identified in turn through feature extraction,and all subgraphs are stitched to realize the detection of road area in the image.Experimental results show that the proposed algorithm is not easily affected by shadows,vehicles and pedestrians,and it has high accuracy.(4)This thesis proposes a road detection algorithm based on fully convolutional networks.Firstly,the algorithm constructs a fully convolutional networks structure model.Secondly,it uses the training set in the KITTI database to train and adjust the network parameters to obtain the training result of the model.Finally,it uses the trained network model to detect the road area in the image.The algorithm has achieved good detection results on the KITTI test set and it is not easily affected by shadows,vehicles,and multi-lane marker lines in the image.