Pedestrian detection in underground mines via parallel feature transfer network

Abstract

Pedestrian detection has been one of the key technologies in computer vision for autonomous driving in underground mines. However, such pedestrian detection is easily affected by complex environmental factors, such as uneven light, dense dust and cable interference. Recently, the problem of pedestrian detection is solved as an object detection task, which has achieved significant advances with the framework of deep neural networks. In this paper, we propose a novel parallel feature transfer network based detector called PftNet that achieves better efficiency than one-stage methods and maintains comparable accuracy of two-stage methods. PftNet consists of two interconnected modules, i.e., the pedestrian identification module and the pedestrian location module. The former aims to roughly adjust the location and size of the anchor box, filter out the negative anchor box, and provide better initialization for the regression. The latter enables PftNet to adapt to different scales and aspect ratios of objects and further improves the regression accuracy. Meanwhile, a feature transfer block compromising gated units is well designed to transmit the pedestrian characteristics between two modules. Extensive experiments on self-annotated underground dataset as well as INRIA and ETH datasets show that PftNet achieves state-of-the-art detection efficiency with high accuracy, which is significant to realizing unmanned driving systems in mines.