Mfmam-Yolo: A Method for Detecting Pole-Like Obstacles in Complex Environment

• DOI: 10.2139/ssrn.4535923
• Publication Date: 2023-08-09T07:48:26Z
• Authors: Lei Cai; Hao Wang; Congling Zhou; Yongqiang Wang; Boyu Liu
• Keywords: FOS: Computer and information sciences; I.4.1; I.2.10; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition; I.4.1; I.2.10

In real-world traffic, there are various uncertainties and complexities in road weather conditions. To solve the problem that feature information of pole-like obstacles complex environments is easily lost, resulting low detection accuracy real-time performance, a multi-scale hybrid attention mechanism algorithm proposed this paper. First, optimal transport function Monge-Kantorovich (MK) incorporated not only to overlapping multiple prediction frames with matching but also MK can be regularized prevent model over-fitting; then, features at different scales up-sampled separately according optimized efficient pyramid. Finally, extraction space channel enhanced based on mechanism, which suppresses irrelevant environment background focuses obstacles. Meanwhile, paper conducts real test experiments variety environments. The experimental results show precision, recall, average precision method 94.7%, 93.1%, 97.4%, respectively, frame rate 400 f/s. This research detect time accurately, further promotes innovation progress field automatic driving.