Detection Of Unexpected Obstacles Based On Stereo Vision

Unexpected obstacles,such as a cargo box dropped by the front vehicle,a toy ball rolling on the ground and so on,present serious potential dangers to autonomous vehicles.Although detection of pedestrians,vehicles,traffic signs and other common objects based on deep learning has achieved great success in recent years,there is little research on detection of unexpected obstacles.Since the shape,color and texture of unexpected obstacles are not well-defined,learning-based object detection algorithms may not be able to effectively detect the obstacles.In this paper,detailed studies on detection of unexpected obstacles are carried out based on binocular images.Detailed works are as follow:First,unexpected obstacle detection based on stereo depth information.In this paper,stereo images are used to obtain depth information and stick pixel representations of obstacles are obtained by ground detection and disparity segmentation.During ground detection,edge detection and Hough detection are used to regress the ground line,and the detection results of unexpected obstacles is improved.Second,unexpected obstacle detection based on image segmentation network.In this paper,the image semantic segmentation network based on convolution neural network is used to extract the features of unexpected obstacles and the surrounding environment.ENet network has been used and training images are cropped vertically to increase the training samples.Hence,the detection rate of unexpected obstacle is further improved while ensuring the network efficiency.Finally,unexpected obstacle detection based on the combination of stereo depth information and image semantic segmentation.In this paper,the energy function optimization algorithm based on the combination of geometric and semantic information is used to combine the detected ground,disparity information and semantic information.The experiments show that the detected obstacle boundaries have been improved.To sum up,this paper studies the detection of unexpected obstacles from three aspects: based on stereo depth information,based on image semantic segmentation and the combination of two kinds of information.The research results are valuable for related study on unexpected obstacles and they are helpful to promote the application of intelligent traffic driving.