Unmanned driving-oriented underground mine pedestrian detection method

Abstract

Pedestrian detection is a key technology for unmanned driving in underground coal mine, which is affected by uneven illumination, complex background, infrared interference, dim light and small and dense targets in images, etc. The existing methods are not ideal for detecting pedestrians in underground mines. In order to solve the above problems, a multi-sensor fusion method for underground mine pedestrian detection is proposed. This method uses a step-by-step multi-characteristic fusion method to fuse the image characteristics collected by the visible light sensor, infrared sensor and depth sensor to obtain richer image characteristics. On the basis of RetinaNet, Dense connection is added to ResNet to form a Dense-ResNet with a hierarchical connected structure, which is able to extract the deep image characteristics from the multi-sensor fusion results and enhance the detection capability of small targets. The experimental results show that multi-sensor fusion images can obtain richer target characteristics compared with a single image, which is beneficial to improve the target detection accuracy. Compared with RetinaNet, Dense-RetinaNet can improve the accuracy of multi-target and small target detection.

Keywords

• underground mine unmanned driving; underground mine pedestrian detection; multi-sensor fusion; characteristic extraction; retinanet; dense connection