An Efficient Lightweight Object Detector for Railway Tunnel Safety Monitoring

Tunnel disaster usually poses a huge threat to trains and passengers, hence the monitoring of the tunnel environment becomes particularly important. In this paper, we aim to detect the potential tunnel disasters from the perspective of computer vision. An efficient lightweight network is proposed to mainly detect pedestrians and trains in tunnel, a lightweight backbone is leveraged to reduce the volume of network parameters and computational costs, while multi-scale feature fusion enhances the spatial and semantic features of various layers. As a foreground mask to the Gaussian Mixture Model, our detector aims to improve the performance of obstacle detection by reducing the false alarm rate. A number of experiments are carried out in the tunnel laboratory. According to experimental results, the detection framework proposed in this paper beats the state-of-the-art detectors in tunnel dataset. Further, the mask generated by the detector significantly decreases the false alarm rate of the Gaussian Mixture Model of obstacle detection, which proves that the framework proposed in this paper is applicable to practical tunnel safety monitoring.