A Faster Neural Network for High-speed Railroad Environment Safety Hazard Segmentation and Risk Evaluation Using UAV Imagery

Environmental safety hazards along the high-speed railroad (e.g. rubbish, plastic tarps, and temporary dwellings) will threaten the safety operation of train. However, existing deep learning-based approaches are constrained by scale-significance differences in safety hazards and limited semantic information, as well as inadequate risk evaluation of hazard instances. This paper proposes a faster neural network named FRESeg for environment safety hazard segmentation and risk evaluation using UAV (Unmanned Aerial Vehicle) imagery. Firstly, a novel cross stage partial connection module integrating partial convolution is developed and assembled into the FRESeg for mining rich feature information while keeping a low latency. Second, an image-to-real-world mapping model taking advantage of key UAV parameters (such as, flight altitude, lens focal length, etc.) is also integrated in the FRESeg for more complete estimation of risk distance and area. Finally, the comprehensive experiment results utilizing the complex UAV high-speed railroad dataset show that proposed method acquires a faster inference speed yet maintains a higher detection rate.