Detection and discrimination of obstacles to vehicle environment under convolutional neural networks

Nowadays driving safety has become an important issue. Obstacles in the driving environments is a main risk of driving. In order to meet the requirements of safety driving, we set up a novel obstacle detection and recognition method based on neural network. In our proposed method, a multi-layer neural network is established to build a 15- layer network that contains the recommended areas for vehicle obstacle and obstacle detection. The convolution neural network (CNN) is trained. To split the region of interest (ROI) completely, on the basis of the ROI algorithm, the maximum variance of the region growing algorithm is used to automatically obtain the threshold value, that also can obtain the segmentation boundary of image. Meanwhile, the morphological operation is used to smooth the boundary of image. By analyzing the accuracy and computation cost of the experimental results, the CNN network in this paper achieves great results in obstacle detection and discrimination.