Human Activity Recognition Using Cascaded Dual Attention CNN and Bi-Directional GRU Framework

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Vision-based human activity recognition (HAR) is crucial for video analytics.
• Existing deep learning methods for HAR often present a trade-off between accuracy and computational efficiency.
• There is a need for HAR models that are both accurate and computationally efficient.

Purpose of the Study:

• To propose a computationally efficient and generic spatial-temporal cascaded framework for HAR.
• To enhance both the accuracy and computational efficiency of human activity recognition models.
• To address the limitations of current HAR methods that prioritize either performance or efficiency.

Main Methods:

• Developed an efficient dual attentional convolutional neural network (DA-CNN) for salient feature extraction using a unified channel-spatial attention mechanism.
• Employed a stacked bi-directional gated recurrent unit (Bi-GRU) for long-term temporal modeling and action recognition.
• Integrated DA-CNN and Bi-GRU into a cascaded framework for spatial-temporal feature exploitation.

Main Results:

• The proposed DA-CNN+Bi-GRU framework demonstrated superior performance over state-of-the-art methods on three public HAR datasets.
• Achieved significant improvements in both model accuracy and inference runtime.
• Showcased an execution time improvement of up to 167x in frames per second compared to contemporary methods.

Conclusions:

• The DA-CNN+Bi-GRU framework effectively balances accuracy and computational efficiency for HAR.
• The proposed model offers a robust solution for real-time video analytics tasks.
• This approach advances the field of vision-based human activity recognition by providing a highly efficient and accurate model.