A lightweight double-channel depthwise separable convolutional neural network for multimodal fusion gait recognition

Area of Science:

• Biometric technology
• Wearable device security
• Human-computer interaction

Background:

• Gait recognition is an emerging biometric technology for wearable device privacy.
• Existing methods face challenges in performance, model size, and robustness.
• Need for efficient and accurate gait recognition in wearable systems.

Purpose of the Study:

• To propose a novel multimodal gait recognition method for wearable devices.
• To improve recognition accuracy while reducing model memory footprint and enhancing robustness.
• To preserve temporal data dependencies using the Gramian angular field (GAF) algorithm.

Main Methods:

• Multimodal fusion of gait cycle data and Gramian angular field (GAF) images.
• Development of a lightweight double-channel depthwise separable convolutional neural network (DC-DSCNN).
• Extraction of gait features using a three-layer depthwise separable convolutional neural network (DSCNN) module and softmax classification.

Main Results:

• The proposed DC-DSCNN algorithm achieved a recognition accuracy of 99.58% on a dataset of 24 subjects.
• The model's memory usage was significantly reduced to only 972 KB.
• Experimental results demonstrate high accuracy and low memory consumption.

Conclusions:

• The proposed method offers a robust and efficient solution for gait recognition using wearable devices.
• The DC-DSCNN model enables lower power consumption and higher real-time performance for gait-based authentication.
• This approach enhances privacy protection for wearable device users through accurate and lightweight biometric identification.