ABSTRACT For autonomous vehicles (AVs) and advanced driver assistance systems (ADAS) in rail transit, highly accurate environmental perception is essential for warning of potential hazards and proactively preventing accidents. However, most existing ADAS solutions rely on monocular or multi‐camera setups with similar focal lengths, which are insufficient for the real‐time detection of both long‐distance and wide‐field obstacles. To address this limitation, we propose a lightweight dual‐path fusion network (LDPFN) that leverages dual‐path cameras with different focal lengths for both road and rail applications. Specifically, a short‐focal camera provides wide coverage for detecting roadside or trackside obstacles, while a long‐focal camera captures long‐range small obstacles. To achieve effective online registration between the two camera streams, we employ an enhanced ORB feature‐point matching strategy. Furthermore, a lightweight MobileNet backbone is adopted to reduce computational and memory overhead, while spatially‐adaptive feature modulation is integrated to dynamically select multi‐scale feature representations for improved image feature extraction. Experimental results show that the proposed method achieves superior accuracy compared to state‐of‐the‐art approaches, with a mean average precision of 0.96 and an inference time of 26.5 ms per image on an RTX 3060 GPU. These results demonstrate that LDPFN effectively balances detection accuracy and real‐time inference speed, offering a practical solution for obstacle detection in AVs and ADAS applications.
Cai et al. (2026) studied this question.