Adaptive weighted imbalance learning with application to abnormal activity recognition

Abnormal activity recognition has been paid much attention in the field of healthcare and related applications, especially for the elderly people's physical and mental health, the high risk of the fall accident and its caused injures have gradually attracted more and more concerns. At present, wearable devices based fall detection technology can effectively and timely monitor the occurrence of fall accidents and help the injured person receive the first aid. However, the built classifiers of traditional approaches for fall detecting and monitoring suffer from a high false-alarm rate though they can reach a relatively high detection accuracy, further they have to face with the imbalance problem because sensor data of abnormal activities are usually rare in the realistic application. To address this challenge, we propose two-stage adaptive weighted extreme learning machine (AWELM) method for eyeglass and watch wearables based fall detecting and monitoring. Experimental results validate and illustrate significant efficiency and effectiveness of the proposed method and show that, our approach firstly achieves a good balance between high detection accuracy and low false-alarm rate based on our two-stage recognition scheme; secondly enables our imbalance learning approach for scarce abnormal activity data by two-stage adaptive weighted method; thirdly provides a light-weight classifier solution to resource constrained wearable devices using extreme learning machine with the fast training speed and good generalization capability, which enables large-scale mHealth applications and especially helps the elderly people to greatly reduce the risk of fall accidents finally.