Abstract:

Aiming at the difficulty of recognizing the obstruction of the human body in public, a method of using Kinect depth data to solve the positive gait recognition in the occluded scene is proposed. First, the image is captured by installing a depth camera at the top of the entrance and exit of the surveillance area, and the image is segmented by the background subtraction method, the RGB color space is normalized to detect and remove shadows to complete the image preprocessing. Then, the periodic changes of the skeleton structure of the lower body area evaluated by the Kinect are extracted from the front view, and the feature sets corresponding to the rear view are extracted from the depth information of the shadow outline. These feature sets retain high-resolution gait action information. Finally, the unidentified frame of a cluttered test sequence is compared with the matched frame of the training sequence to complete the final recognition. Experiments show that this method is computationally efficient and achieves satisfactory results at different levels of occlusion.

Key words: Frontal gait recognition, Kinect depth data, occluded scenes, RGB, background subtraction method

摘要： 针对公共场合人体目标易被遮挡而造成的识别困难，提出了一种利用Kinect深度数据来解决遮挡场景中正面步态识别的方法。首先，通过在监控区域出入口处的顶部安装深度相机来采集图像，并用背景减除法进行图像分割，再归一化RGB颜色空间来检测和去除阴影完成图像的预处理。然后从前视图提取由Kinect评估的下半身区域骨架结构的周期性变化，从阴影轮廓的深度信息提取后视图对应的特征集，这些特征集保留了高分辨率的步态动作信息。最后，将一个杂乱的测试序列的未遮挡帧与训练序列的匹配帧进行比较，完成最终的识别。实验证明，该方法在计算上有效，并在不同遮挡程度下取得了令人满意的结果。

关键词: 正面步态识别, Kinect深度数据, 遮挡场景, RGB, 背景减除法