PROJECT TITLE :

Learning Compact Features for Human Activity Recognition Via Probabilistic First-Take-All

ABSTRACT:

With the rise in popularity of mobile sensor technologies, smart wearable devices provide a once-in-a-lifetime opportunity to solve the difficult challenge of human activity recognition (HAR) by learning expressive representations from multidimensional everyday sensor information. This motivates us to create a novel algorithm that can be used in both camera-based and sensor-based HAR systems. Despite reports of competitive classification accuracy, present approaches frequently struggle to identify visually similar activities built of activity patterns in distinct temporal sequences. We present a unique probabilistic approach for compactly encoding temporal ordering of activity patterns for HAR in this study. The algorithm, in particular, learns an ideal set of latent patterns whose temporal structures are important in distinguishing various human activities. Then, using a novel probabilistic First-Take-All (pFTA) approach, compact features are generated from the orders of these latent patterns to encode the entire sequence, and the Hamming distance between compact features is used to efficiently measure the temporal structural similarity between different sequences. Experiments on three publicly available HAR datasets show that the suggested pFTA technique may compete in terms of accuracy and efficiency.