Geometrically Consistent Pedestrian Trajectory Extraction for Gait Recognition
Geometrically Consistent Pedestrian Trajectory Extraction for Gait Recognition
复制标题
DOI：
10.1109/btas.2018.8698559
复制
发表时间：
2018-10
期刊：
影响因子：
--
通讯作者：
Yasushi Makihara;Gakuto Ogi;Y. Yagi
中科院分区：
文献类型：
--
作者：
Yasushi Makihara;Gakuto Ogi;Y. Yagi
在步态识别领域，基于轮廓的步态表示方法，如步态能量图，在过去的十年中得到了广泛的应用。为了获得高质量的步态特征，必须获得对齐良好的轮廓序列，然而，这对于不完美的轮廓以及由于透视投影而导致的尺度和位置变化并不一定容易。因此，我们提出了一种面向步态识别的行人轨迹提取方法，即。例如，每个行人的边界框序列。更具体地说，我们首先开发了一个交互式的工具来获得摄像机校准参数的目标场景几何没有现场的工作量。然后，我们引入了一个几何约束，以更好地保持框架之间的边界框的一致性w.r.t.行人的高度和脚底指向地平面。其次，我们反复应用分析动态规划（DP）来找到多个行人在地平面上的轨迹，其中基于语义分割结果和颜色直方图相似性来计算数据和过渡分数。此外，由于DP只考虑相邻帧之间的平滑性，我们通过分段线性轨迹来近似轨迹，使其更具全局平滑性。实验结果表明，该方法使我们能够更好地对齐步态特征，从而提高步态识别的准确性。
In the gait recognition community, silhouette-based gait representations such as gait energy image have been widely employed for the last decade. In order to obtain good quality of gait features, it is essential to get a well aligned silhouette sequence, which is, however, not necessarily easy with imperfect silhouettes and also with scale and position changes due to perspective projection. We therefore propose a gait recognition-oriented approach to pedestrian trajectory extraction, i. e., bounding box sequence for each pedestrian. More specifically, we firstly developed an inter-active tool to get camera calibration parameters for a target scene geometry without on-site workload. We then introduce a geometric constraint to better keep the consistency of bounding boxes among frames w.r.t. the pedestrian ’s height and foot bottom points on the ground plane. Sub-sequently, we apply analytical dynamic programing (DP) repeatedly to find multiple pedestrian ’s trajectories on the ground plane, where data and transition scores are computed based on semantic segmentation results and color histogram similarity. Moreover, since DP just considers the smoothness between adjacent frames, we approximate the trajectory by a piece-wise linear trajectory to make it more globally smooth. Experimental results show that the proposed method enables us to make better aligned gait features and consequently improves gait recognition accuracy.