Continuous Human Action Recognition Using Depth-MHI-HOG and a Spotter Model

Abstract

In this paper, we propose a new method for spotting and recognizing continuous human actions using a vision sensor. The method is comprised of depth-MHI-HOG (DMH), action modeling, action spotting, and recognition. First, to effectively separate the foreground from background, we propose a method called DMH. It includes a standard structure for segmenting images and extracting features by using depth information, MHI, and HOG. Second, action modeling is performed to model various actions using extracted features. The modeling of actions is performed by creating sequences of actions through k-means clustering; these sequences constitute HMM input. Third, a method of action spotting is proposed to filter meaningless actions from continuous actions and to identify precise start and end points of actions. By employing the spotter model, the proposed method improves action recognition performance. Finally, the proposed method recognizes actions based on start and end points. We evaluate recognition performance by employing the proposed method to obtain and compare probabilities by applying input sequences in action models and the spotter model. Through various experiments, we demonstrate that the proposed method is efficient for recognizing continuous human actions in real environments.

Depth Image-Based Human Action Recognition Using Convolution Neural Network and Spatio-Temporal Templates

“…그러나 몸 전체를 인식하는 시스템은 특정 부분을 인식하는 시스 템 보다 많은 문제점이 생긴다. 그래서 몸 전체를 인식에 대한 연구와 개발이 인식 성능을 개선하기 위해 진행되고 있다 [5,6].…”

An Online Continuous Human Action Recognition Algorithm Based on the Kinect Sensor

“…Compared with the CHAR methods [ 18 , 19 ] which need to detect the start and end points of human actions beforehand, the proposed algorithm does not detect the start and end points of human actions explicitly. The proposed algorithm divides feature sequences into pose feature segments and motion feature segments, and the potential-difference-based segmentation method ensures that almost all the boundaries of human actions can be conformed with boundaries of the first and last segments within each human action’s feature sequence.…”

Robust Indoor Human Activity Recognition Using Wireless Signals

“…Many studies have been done in recent years [1,2,3,4,5], however, robust action recognition is still a challenging problem due to the following issues: (a) Body parts or big size obstacles may cause partial occlusions; (b) An action, observed from different viewpoints, has different appearances; (c) Clothing, especially long skirts, may lead to apparent anthropometric differences; (d) The start-time and end-time points of an action are sometimes hard to detect accurately; (e) Dynamic backgrounds may make it difficult to locate and observe actions; (f) Smoke-filled, dim, or dark rooms may make it hard to observe actions; (g) People may feel uncomfortable with a camera overhead, especially in a bathroom. To tackle any one of these problems is an arduous task, and they are likely to appear at the same time in practical applications.…”