Gait recognition from videos is a very important task for surveillance video analysis. Although a number of studies have explored gait recognition models, they lack clarity in the gait contour segmentation, which is an important but difficult step for automatic gait recognition. Most of the gait recognition algorithms use manually segmented gait contours, which is not available in real situations and not suited for real-time video processing applications. To date, there are very little research directly investigating automatic pedestrian gait contour segmentation. Current state-of-the-art instance segmentation methods fail to accurately describe the contour of whole pedestrian body and often deviate from the accurate boundaries, especially for the contour between two legs, which is the essential information for gait recognition. This paper presents a novel gait contour automatic segmentation model (GCA-Net) for gait recognition in videos. To improve the segmentation and edge fitting accuracy, we firstly use the dilated convolutions in the residual block to enhance the feature representative ability of the ResNet backbone, and then an edge detection module is added to the model which can make the predicted gait contour closer to the actual boundaries and therefore improve the edge fitting result. The experiment results show the effectiveness of the proposed method. The edge detection module can increase the performance by 5.4% and the residual block with dilated convolution can further increase the performance by 0.4%. More important, the proposed model can be directly integrated into existing gait recognition methods and automate video gait recognition.

视频步态识别是监控视频分析的一项非常重要的任务。尽管有大量研究探索了步态识别模型，但它们在步态轮廓分割方面缺乏清晰度，这是自动步态识别的重要但困难的一步。大多数步态识别算法使用手动分割的步态轮廓，这在实际情况下不可用，也不适合实时视频处理应用程序。迄今为止，很少有研究直接研究自动行人步态轮廓分割。当前最先进的实例分割方法无法准确描述整个行人身体的轮廓，并且经常偏离准确的边界，尤其是两条腿之间的轮廓，这是步态识别的基本信息。本文提出了一种用于视频步态识别的新型步态轮廓自动分割模型（GCA-Net）。为了提高分割和边缘拟合的精度，我们首先在残差块中使用扩张卷积来增强 ResNet 主干的特征代表能力，然后在模型中加入边缘检测模块，使预测的步态轮廓更接近于实际边界，从而改善边缘拟合结果。实验结果表明了所提出方法的有效性。边缘检测模块可以将性能提高 5.4%，带有空洞卷积的残差块可以进一步提高 0.4% 的性能。更重要的是，所提出的模型可以直接集成到现有的步态识别方法中，实现视频步态识别的自动化。