Human Activity Recognition using Attribute-Based Neural Networks and Context Information

Maria J. Danford

Accurate human activity recognition (HAR) is critical in programs these types of as health care and sporting activities. Deep neural networks are effectively utilised for the undertaking on the other hand, they do not use prior information offered in extremely structured domains.

For case in point, guide function procedures are structured into procedure actions, and facts about them is offered directly from external resources.

Gesture recognition. Picture credit score: Comixboy by way of Wikimedia, GNU Totally free Documentation License.

A the latest paper reveals how context facts can be integrated into a HAR method. A deep neural network extracts motion descriptors, like posture, from raw wearable-sensor data. A classifier estimates activity lessons from the characteristics and optionally from the context facts.

The proposed architecture lets integrating context facts with out re-coaching the network. Empirical evaluation reveals that the product achieves elevated effectiveness, in comparison to the point out-of-the-art, even when no context facts is offered.

We consider human activity recognition (HAR) from wearable sensor data in guide-function procedures, like warehouse purchase-selecting. This kind of structured domains can usually be partitioned into distinct procedure actions, e.g., packaging or transporting. Each and every procedure phase can have a distinct prior distribution over activity lessons, e.g., standing or going for walks, and distinct method dynamics. Right here, we exhibit how these types of context facts can be integrated systematically into a deep neural network-based HAR method. Specifically, we propose a hybrid architecture that combines a deep neural network-that estimates substantial-level motion descriptors, characteristics, from the raw-sensor data-and a shallow classifier, which predicts activity lessons from the approximated characteristics and (optional) context facts, like the presently executed procedure phase. We empirically exhibit that our proposed architecture improves HAR effectiveness, in comparison to point out-of-the-art methods. Moreover, we exhibit that HAR effectiveness can be additional elevated when facts about procedure actions is incorporated, even when that facts is only partially proper.

Investigate paper: Lüdtke, S., Moya Rueda, F., Ahmed, W., Fink, G. A., and Kirste, T., “Human Action Recognition working with Attribute-Dependent Neural Networks and Context Information”, 2021. Hyperlink: muscles/2111.04564

Next Post

MedPerf: Open Benchmarking Platform for Medical Artificial Intelligence using Federated Evaluation

AI could make healthcare treatment a lot more successful, personalized, and cost-effective when implemented at scale. In a study posting just lately released on and titled “MedPerf: Open Benchmarking Platform for Healthcare Synthetic Intelligence making use of Federated Evaluation”, a staff of researchers have proposed MedPerf platform for the […]

Subscribe US Now