Real-time Obstacle Detection Over Rails Using Deep Convolutional Neural Network

In vision-based environment perceptive system of urban rail transit, cameras installed in the front of the train can assist people to identify obstacles on rails. Varying environment makes small targets like pedestrians and bags hard to identify, thus the real-time and accuracy performance of the detection need to be improved. Inspired by the achievements of Single Shot Multibox Detection (SSD) successfully applied on images recognition in recent years, we presented an obstacle detection algorithm which consists of two steps: main network and feature fusion. In the first part, the input image is converted into multi-scale feature maps based on the Residual Neural Network. Next, a series of convolution layers are added to extract features, and the network outputs a confidence score and bounding boxes for possible obstacles. Experiments showed that the proposed method can detect obstacles in various environment. Compared with traditional object detection algorithms and other deep learning algorithms, it runs at a faster detection speed (26 frames per second, FPS) and higher detection accuracy (91.61% mean average precision, mAP) on our self-made dataset.