Human activity recognition using a multi-branched CNN-BiLSTM-BiGRU model

Human behaviour analysis, human-computer interaction, and pervasive computing are three areas where human activity recognition has recently attracted a lot of attention. Recent advances in deep learning have made it possible to accurately predict a variety of human actions by handling time -series data from wearable sensors and mobile devices quickly. Despite their remarkable achievements in recognising activities, DL -based algorithms continue to face hurdles in successfully processing sequential data. Ongoing difficulties include complex feature extraction, dealing with data imbalances, and other related concerns. Furthermore, manual feature engineering techniques are significantly used in the majority of HAR approaches. In order to recognise human behaviours using wearable sensors, this research introduces a powerful classification technique. The strategy combines a bidirectional long -short-term memory, a convolutional neural network, and a bidirectional gated recurrent unit. This system efficiently and successfully derives critical insights from unprocessed sensor data. The model gains the capacity to recognise both short-term patterns and long-term associations in sequential data by merging CNN, BiGRU, and BiLSTM components. The model can capture a wide range of temporal local connections since the feature extraction process is sped up by the addition of multiple filter sizes. The algorithm's performance is assessed using common datasets, particularly the WISDM dataset, which shows a remarkable accuracy of 99.33% and an F1 -score of 73.20% for the hybrid model CCBB. The proposed model is proven to be superior to competing approaches by experimental findings.