A comprehensive survey on deep learning methods in human activity recognition (HAR) is presented, highlighting the latest advancements in both classical machine learning and deep learning techniques. The survey addresses the challenges of navigating the vast HAR literature and introduces a novel methodology using large language models (LLMs) to efficiently filter and identify relevant academic papers. It also provides a taxonomy of HAR methods, distinguishing between sensor-based and vision-based approaches, and offers a structured overview of the field. The survey covers a wide range of input modalities, including accelerometer data, video sequences, and audio signals, and discusses various datasets and algorithms used in HAR research. It emphasizes the importance of preprocessing steps, feature extraction, and the role of deep learning in modeling complex temporal relationships. The survey also identifies key challenges in HAR, such as the integration of multi-sensor data, the high cost of labeled data, and the need to account for simultaneous activities. The methodology includes data collection from academic repositories, natural language processing for keyword extraction and paper filtering, and a detailed taxonomy of HAR methods based on cost, approach, performance, datasets, and supervision. The survey concludes with a discussion of future research directions, emphasizing the potential of vision-based and deep learning methods in HAR.A comprehensive survey on deep learning methods in human activity recognition (HAR) is presented, highlighting the latest advancements in both classical machine learning and deep learning techniques. The survey addresses the challenges of navigating the vast HAR literature and introduces a novel methodology using large language models (LLMs) to efficiently filter and identify relevant academic papers. It also provides a taxonomy of HAR methods, distinguishing between sensor-based and vision-based approaches, and offers a structured overview of the field. The survey covers a wide range of input modalities, including accelerometer data, video sequences, and audio signals, and discusses various datasets and algorithms used in HAR research. It emphasizes the importance of preprocessing steps, feature extraction, and the role of deep learning in modeling complex temporal relationships. The survey also identifies key challenges in HAR, such as the integration of multi-sensor data, the high cost of labeled data, and the need to account for simultaneous activities. The methodology includes data collection from academic repositories, natural language processing for keyword extraction and paper filtering, and a detailed taxonomy of HAR methods based on cost, approach, performance, datasets, and supervision. The survey concludes with a discussion of future research directions, emphasizing the potential of vision-based and deep learning methods in HAR.