Soft assignment and multiple keypoint analysis-based pedestrian counting method

Abstract

Pedestrian counting in videos is an active computer vision research topic that has wide ranging application. Existing pedestrian counting methods predominantly use features extracted from the foreground following subtraction of the background. However, accurately locating the foreground in real environments is difficult, and background subtraction is computationally expensive. The keypoint approach, which counts pedestrians without background subtraction, is limited owing to lack of sufficient features and no consideration for stationary pedestrians. This letter proposes an accurate keypoint-based pedestrian counting method. As no single keypoint detector can yield optimal counting results under all conditions, such as image resolution, frame rate, and illumination, we combine complementary keypoint detectors to enrich the features and thereby enhance pedestrian counting results. In addition, the proposed method considers stationary pedestrians by analyzing static keypoints information. Information loss during vector quantization is also reduced by applying soft assignment during feature extraction. The results of experiments conducted on public databases indicate that the proposed method outperforms the state-of-the-art methods on realistic outdoor and indoor public datasets.