Efficient Human Activity Recognition Solving the Confusing Activities via Deep Ensemble Learning

Background

Human activity recognition (HAR), identifying the actions of individual based on a set of observation, is widely applied to pedestrian indoor tracking, healthcare, Intelligent city, etc. Vision-based recognition algorithms suffer from factors such as lighting condition, clothing color, and image background. Therefore, this paper, leveraging the build-in sensor of the smartphone, proposes a sensor-based HAR model.

Contribution

The proposed model utilizes CNN to perform the HAR task. A CNN-7 block is used to identify seven activities. In terms of confusing human activities, a binary-classifier CNN-2, in the ensemble learning way, helps the model to determine the actual human activity through weighted voting. The contributions can be summarized into:

• Abundant and extensive date collection: The data in this paper contains three motion sensors readings collected by 100 participants. There are four different placements of the smartphone and seven typical daily human activities. The data set not only guarantees the effective training on neural network, but makes the recognition task challenging.
• Accurate and robust recognition: Avoiding handcrafted features engineering, the average accuracy of proposed ensemble model can achieve up to 96.11%. The devised voting mechanism improves the classification accuracy among two confusing activities (walking and going upstairs) further.