Effects of sliding window variation in the performance of acceleration-based human activity recognition using deep learning models

• DOI: 10.7717/peerj-cs.1052
• Publication Date: 2022-08-08T08:31:53Z
• Authors: Milagros Jaén-Vargas; Karla Miriam Reye...; Francisco Fernandes; Sérgio Barroso Go...; Miguel Tavares Silva; Daniel Simões Lopes; José Javier Serra...
• Keywords: Sliding window protocol; F1 score; Feature (linguistics)

Deep learning (DL) models are very useful for human activity recognition (HAR); these methods present better accuracy HAR when compared to traditional, among other advantages. DL learns from unlabeled data and extracts features raw data, as the case of time-series acceleration. Sliding windows is a feature extraction technique. When used preprocessing it provides an improvement in accuracy, latency, cost processing. The time can be beneficial especially if window size small, but how small this keep good accuracy? objective research was analyze performance four models: simple deep neural network (DNN); convolutional (CNN); long short-term memory (LSTM); hybrid model (CNN-LSTM), variating sliding using fixed overlapped identify optimal HAR. We compare effects two acceleration sources': wearable inertial measurement unit sensors (IMU) motion caption systems (MOCAP). Moreover, short sizes 5, 10, 15, 20, 25 frames ones 50, 75, 100, 200 were compared. fed acquired experimental conditions three activities: walking, sit-to-stand, squatting. Results show that most 20-25 (0.20-0.25s) both sources, providing 99,07% F1-score 87,08% (CNN-LSTM) 98,8% 82,80% MOCAP data; similar accurate results obtained with LSTM model. There almost no difference larger (100, 200). However, smaller decrease F1-score. In regard inference time, 20 preprocessed around 4x (LSTM) 2x times faster than 100 frames.