Human activity recognitionHuman Activity Recognition (HAR) (HAR) has gained increasing relevance in health care, especially with the rise of wearable sensors and mobile health technologies. By capturing and analyzing patterns in human motion, HARHuman Activity Recognition (HAR) systems offer a noninvasive method to monitor daily activities and detect irregularities that may signal the onset or progression of disease. HAR research works and system prototypes developed so far mainly utilize accelerometerAccelerometer and gyroscopeGyroscope data. Apart from these multimodal sensor data is also being efficiently used in deep learningDeep Learning (DL)-based HARHuman Activity Recognition (HAR) models. Other sensor data utilized is of ambient sensors such as RFID, smart homes, video surveillance, or of vision-based sensors such as camera sensor and depth sensors. Deep LearningDeep Learning (DL) (DL) in Human Activity Recognition (HAR)Human Activity Recognition (HAR) utilizing multimodal sensory data is getting immensely popularized in compared to machine learningMachine Learning (ML) because that Deep learningDeep Learning (DL) can do automatic feature extraction from raw sensory data for more accurate pattern identification. Deep learning has revolutionary impact on human activity recognitionHuman Activity Recognition (HAR) by automatic feature learning advantage over handcrafted feature extractionFeature extraction of machine learningMachine Learning (ML). Also deep learningDeep Learning (DL) can accurately identify human activities through heterogeneous multimodal data ranging from motion sensory data to image to video data, etc. There are many popularly used deep learningDeep Learning (DL) models, and few of them are convolutional neural networksConvolutional Neural Networks (CNN) (CNNs), recurrent neural networks (RNNs) specifically LSTMsLong Short-Term Memory (LSTM)/GRUs, hybrid models, e.g., CNN and LSTM combination, graph neural networks (GNNs), etc. CNN efficiently captures spatial dependencies in sensor-based signal, and RNN efficiently models temporal dependencies in sequential activity data. Hybrid models combine spatial–temporal feature extractionFeature extraction techniques to achieve improved performance. For skeleton-based human activity recognitionHuman Activity Recognition (HAR), GNNs are being popularly used analyzing graph-structured pose data.