基于关键区域的二值化场景特征快速提取方法

Key region identification with binary feature extraction for high accurate scene matching

近年来,驾驶辅助系统中基于视频信息的车辆定位技术受到广泛关注。针对轻轨系统高精度场景匹配中场景相似度过高导致定位困难的问题,提出了一种关键区域及二值化特征提取方法。该方法以离线处理的方式在高相似度的参考序列帧内提取具有显著性信息的关键区域,并在这些区域中生成二值化特征描述符以提高实时场景匹配的速度与准确率。在香港轻轨数据集以及公开的Nordland数据集中,相对于局部场景特征,基于提出的关键区域特征的场景匹配方法错误偏差下降31.43%,同时节约了94.22%的匹配时间;与Seq SLAM场景跟踪算法相比,在不显著增加运行时间的前提下,基于关键区域二值化场景特征的场景跟踪正确率提高了9.84%。实验结果表明,提出的关键区域以及二值化特征提取方法在降低了场景匹配计算时间的同时,提高了匹配精确度。

As one of the most important component in advanced driver assistance systems, the visual-based localization techniques have been widely studied in recent years. Dealing with the problems of high accurate scene matching caused by the extremely similar scene appearances in the high frame rate reference sequence, a new and powerful method of key region and binary feature extraction is proposed in this paper. The key regions with discriminative information are extracted from the similar reference frames. The binary patterns of these key regions are generated to reduce the computational complexity of online scene matching procedure. The proposed method is evaluated on the Hong Kong light railway dataset and Nordland dataset. With the proposed key regions, the error and the computation time of scene matching are reduced by 31.43% and 94.22%. The precision rate of scene tracking with proposed method is 9.84% higher than that of Seq SLAM.The experimental results show that the proposed method has high performance.