LRF Large-Kernel Attention Convolution Network for Wearable Activity Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current human activity recognition methods based on large receptive fields face challenges in deploying deep neural networks on mobile devices due to increased computational burdens and the need for more calculation parameters, limiting their application in wearable devices with limited computing power.
Innovation Solution
The proposed method employs a large-kernel attention convolution network with a three-layer structure, including local depth convolution, long-distance depth expansion, and 1×1 ordinary convolution layers, along with a self-attention mechanism and layer normalization, to enhance feature extraction and recognition accuracy while reducing computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of network layers is increased to expand the effective receptive field, then activity recognition performance is improved, but computational burden increases significantly
Solution Approach 1:
The patent segments the large-kernel convolution operation into multiple small-kernel convolutions applied in sequence. Specifically, a large kernel of size k×k is decomposed into multiple smaller kernels that are convolved sequentially, achieving the same receptive field expansion without the quadratic increase in computational complexity that would result from directly applying a single large kernel.
Solution Approach 2:
The patent transitions from using a single large kernel in one dimension to using multiple small kernels stacked in the depth dimension. This dimensional transformation allows the network to achieve equivalent receptive field coverage while distributing computational load across multiple lightweight operations rather than one heavy operation.
2Measurement precision
If the size of convolution kernel is increased to expand the effective receptive field, then activity recognition performance is improved, but the number of calculation parameters increases
Solution Approach 1:
The patent segments the large-kernel convolution into multiple small-kernel convolutions. By dividing a large kernel of size k×k into multiple smaller kernels (e.g., three kernels of size k/3×k/3), the total number of parameters is significantly reduced while maintaining the same effective receptive field, since parameters scale with the area of the kernel rather than its linear dimension.
Solution Approach 2:
The patent changes the parameter configuration from a single large kernel with many parameters to multiple small kernels with fewer parameters each. This parameter transformation maintains the mathematical equivalence of the receptive field while reducing the total parameter count, making the model more suitable for deployment on resource-constrained wearable devices.
Data Source
AI summary
The disclosure relates to an activity recognition method of a large receptive field (LRF) large-kernel attention convolution network based on a large receptive field, the method includes the following steps: collecting an action signal, carrying out a preprocessing and a data partition on the action signal to obtain a data set; and training an LRF large-kernel attention convolution network model based on the data set, and introducing a trained LRF large-kernel attention convolution network model into a mobile wearable recognition device for human posture recognition.