Human Behavior Recognition in Outdoor Sports Based on the Local Error Model and Convolutional Neural Network

Area of Science:

• Computer Vision
• Machine Learning
• Embedded Systems

Background:

• Human action recognition is crucial for computer vision applications.
• Deep convolutional neural networks offer high performance but face memory constraints on wearable devices.
• Traditional training methods lead to low memory utilization and backhaul locking issues.

Purpose of the Study:

• To design a novel local error convolutional neural network model for human motion recognition.
• To address the memory limitations of deep convolutional neural networks on embedded wearable devices.
• To improve memory utilization efficiency and overcome backhaul locking problems.

Main Methods:

• Developed a local error convolutional neural network model.
• Implemented layer-by-layer training using local errors.
• Enabled independent parameter training without relying on adjacent layer gradients.

Main Results:

• The local error approach allows for early release of memory used by hidden layer parameters.
• Successfully avoided the backhaul locking problem inherent in traditional methods.
• Significantly improved memory utilization for convolutional neural networks on wearable sensor devices.

Conclusions:

• The proposed local error convolutional neural network is effective for human action recognition on resource-constrained devices.
• This method enhances the feasibility of deploying advanced deep learning models on wearable technology.
• The study offers a viable solution for efficient deep learning on embedded systems.