Novel Deep Learning Network for Gait Recognition Using Multimodal Inertial Sensors.

Area of Science:

• Human-Computer Interaction
• Biomedical Engineering
• Machine Learning

Background:

• Existing gait recognition methods using Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks suffer from information loss (spatial for CNN, temporal for LSTM).
• CNNs capture waveform characteristics but miss crucial time-series gait data, while LSTMs struggle with long sequences, leading to performance degradation.

Purpose of the Study:

• To propose a novel Sequential Convolutional LSTM (SConvLSTM) network for robust gait recognition.
• To automatically extract gait features from multimodal wearable inertial sensor data without manual feature engineering.

Main Methods:

• The SConvLSTM network integrates 1D-CNN for high-dimensional feature extraction and dimension compression, followed by a bidirectional LSTM network for enhanced time-series feature extraction.
• The method processes fixed-length data frames, eliminating the need for gait cycle detection and its associated errors.
• Raw acceleration and gyroscope signals from wearable inertial sensors serve as input data.

Main Results:

• SConvLSTM demonstrated superior performance compared to state-of-the-art methods on three public benchmark datasets: UCI-HAR, HuGaDB, and WISDM.
• The proposed architecture effectively retains time-series features while compressing feature vector length.

Conclusions:

• The SConvLSTM network offers an effective solution for gait recognition by synergistically combining CNN and bidirectional LSTM.
• This approach mitigates the limitations of traditional CNN and LSTM methods, achieving high accuracy without manual feature design or gait cycle detection.