Deep vision-based real-time hand gesture recognition: a review

Hand gesture recognition is an approach to comprehending human body language, applied in various fields such as human-computer interaction. However, some issues remain in edge blurring generated by complex backgrounds, rotation inaccuracy induced by fast movement, and delay caused by computing cost. Recently, the emergence of deep learning has ameliorated these issues, convolution neural network (CNN) enhanced edge clarity, long-short term memory (LSTM) improved rotation accuracy, and attention mechanism optimized response time. In this context, this review starts with the deep learning models, specifically CNN, LSTM, and attention mechanisms, which are compared and discussed from the utilization rate of each, their contribution to improving accuracy or efficiency, and their role in the recognition stage, like feature extraction. Furthermore, to evaluate the performance of these deep learning models, the evaluation metrics, datasets, and ablation studies are analyzed and discussed. The choice of evaluation metrics and dataset is critical since different tasks require different evaluation parameters, and the model learns more patterns and features from diverse data. Therefore, the evaluation metrics are categorized into accuracy and efficiency. The datasets are analyzed from self-created to public datasets. The ablation study is summarized in four aspects: similar underlying models, integrating specific models, pre-processing, others. Finally, the existing research gaps and further research on accuracy, efficiency, application range, and environmental adaptation are discussed.