Human Activity Recognition And Localization For Weakly Labeled Sensor Data

Human activity recognition based on wearable devices is an active and challenging research area that has been explored in many applications such as healthcare,smart surveillance and security detection.With the development of the Internet of Things and the increase in computing power,various sensors are embedded in smart devices to collect the physiological state of the user.Based on the motion data information provided by the user,the computer system is made to assist the user to perform some specific tasks.Some existing research solutions rely on supervised learning approaches that require large amounts of labeled training data,however,precisely labeling the start and end positions of activities is a laborious task.It is worth noting that people usually do not maintain a fixed pattern of movement all the time in their daily lives,but a combination of multiple movement states.Weakly labeled data is easier to collect than strongly labeled data,so it is an interesting direction how to utilize weakly labeled sensor data in human activity recognition systems.The attention mechanism in deep learning mimics human visual processing by selectively focusing attention on regions of interest and ignoring other unimportant regions.This property makes the attention mechanism suitable for processing and solving human activity recognition in weakly supervised mode.In this paper,we combine the advantages of deep learning and attention mechanism to investigate human activity recognition and localization in weakly supervised mode,and the main work accomplished is as follows.(1)Introduce the weakly labeled sensor data used in the experiments of this paper,and briefly describe the evaluation metrics used for human activity recognition and localization.(2)To address the problem of human activity recognition and localization in weakly labeled sensor data,this paper proposes a residual attention mechanism that dynamically computes weights at each time frame and uses the attention weights to detect which frames are more important for target prediction and improve the interpretability of the model.Firstly,the structure of the model is introduced from the principle,and secondly,the feasibility of the attention mechanism is verified by visualizing the weights.Finally,experiments on the weakly labeled sensor dataset show that the activity recognition model based on the residual attention mechanism significantly improves the accuracy of activity classification and identifies the location where the target activity occurs compared to the baseline model.(3)Considering the weakly labeled sensor data as a temporally varying signal,this paper proposes a conditional attention mechanism that enables the model to place more attention on the discriminative subsegments containing the target activity by weighing the similarity between the local features in the convolutional neural network and the conditional global features in the long and short-term memory network to identify and localize the activity without relying on the location information.