Research On Road Detection Methods Based On Multi-feature Fusion In Vehicle-mounted Binocular Scene

In recent years,the road detection problem,as key technology in intelligent assisted driving system,has become a research hotspot in the field of computer vision.In order to better distinguish between road and non-road areas,monocular cameras,LIDAR,binocular cameras and other sensors are put into use.Among all sensors,binocular cameras,which not only are affordable but also provide color and space information of the current scene,are widely used.Therefore,the research object of this paper is the detection of road in vehicle-mounted binocular scene.When we train models based on traditional machine learning methods(manually designed feature +classifier),the quantity of samples demanded is small and requirements for hardware are relatively low.However,when detection models are trained based on the deep convolutional network,a large number of training samples are needed and big enough memory is required.In the case of sufficient samples,the expressive ability of features learned based on the deep convolution network is stronger than that of features manually designed.According to the different requirements,this paper discusses the road detection method based on manually designed feature and the method based on deep convolution network respectively.The main work and innovation of this paper are as follows:(1)To address the problem that a single feature is difficult to describe the road comprehensively,in this paper,we propose a multi-feature fusion based road detection algorithm.According to the characteristics of the road,we use the Gabor texture feature and kernel descriptors which own low loss and high expression for road.In addition,combining these features with 2D features relative to color,gradient and 3D features depicting angle and spatial position,we obtain a high discriminant road feature.The algorithm extracts the corresponding 2D and 3D features within the range of the super pixel and selects the appropriate features for fusion.And then we use a three-layer artificial neural network to help the features better fuse and to do classification.Finally we use a fully connected conditional random field to post process the classified results,as much as possible to eliminate unreliable classification.Experimental results on the KITTI ROAD dataset show that our method outperforms most manually designed feature based methods.(2)Aiming at the problem that the manually designed feature based road detection method has poor performance with complex background,in this paper,we propose an automatic learning and fusion method for color feature and disparity feature based on deep convolutional network.In order to overcome the difficulty of identifying the road and non-road areas in the RGB information in some scenarios,we apply the DeepLab method to the disparity map and fuse it with the RGB information for better detection.The proposed fusion strategy includes fusion based on input data,fusion based on convolution network parameter sharing,and fusion based on feature map addition.The different fusion strategies are evaluated on the KITTI ROAD dataset to select the most effective method,and the effectiveness of the algorithm is verified by experiments.