Continuous Human Action Recognition in Ambient Assisted Living Scenarios

Abstract

Abstract. Ambient assisted living technologies and services make it possible to help elderly and impaired people and increase their personal autonomy. Specifically, vision-based approaches enable the recognition of human behaviour, which in turn allows to build valuable services upon. However, a main constraint is that these have to be able to work online and in real time. In this work, a human action recognition method based on a bag-of-key-poses model and sequence alignment is extended to support continuous human action recognition. The detection of action zones is proposed to locate the most discriminative segments of an action. For the recognition, a method based on a sliding and growing window approach is presented. Furthermore, an evaluation scheme particularly designed for ambient assisted living scenarios is introduced. Experimental results on two publicly available datasets are provided. These show that the proposed action zones lead to a significant improvement and allow real-time processing.

Hierarchical transfer learning for online recognition of compound actions

“…A common adaptation of existing approaches is to use a sliding window and classify the current frame based on the recent temporal history. This enables continuous recognition of multiple actions in real world scenarios such as monitoring elderly patients at home [14]. However, there is an additional requirement in NUI applications to detect actions with low latency so the system can provide an appropriate response to the user with no apparent delay.…”

“…The hierarchical peak key pose for the action is the key pose that has the maximum number of matches, as summarised in algorithm 2. Some existing methods for online action recognition detect the action as a single point in time [9], [17] whereas others incorporate the duration of the action [14], [34]. The duration of the action is important for subsequently detecting interactions between multiple players in a sports game [4] and illustrated by Figure 12.…”

“…Similar to [14], [34] we introduce a threshold for action detection but instead of specifically learning a threshold for each action we learn a single threshold for all actions.…”

“…Key poses remove redundant information to improve classification accuracy and reduce the computational latency of action detection [14], [17]. To generate key poses we follow the method proposed in [31] that uses the training set = { , ′ } ( =1… ) to learn a Radial Basis Function Network (RBFN) that represents the mapping between the embedded and the high dimensional space [31].…”

“…Peak key poses [14] were limited to detecting a single temporal point so we introduce a threshold to incorporate the duration of the peak. Similar to [14], [34] we introduce a threshold for action detection but instead of specifically learning a threshold for each action we learn a single threshold for all actions.…”