Hybrid VGG16-GRU Network for Robust and Real-Time Hand Gesture Recognition

Abstract

Through hand gesture detection, wearables and cars are also bringing more organic and contact-free human-machine interaction to intelligent manufacturing. By merging GRU for temporal sequence modelling and VGG16 for spatial feature extraction, a novel hybrid deep neural network has been developed to overcome the issues of low accuracy, high latency, and lack of tolerance to changes in the actual world. The suggested framework uses the following general pre-processing techniques: adaptive region cropping, strategic data augmentation, histogram-based normalisation, etc. Numerous research has gathered over 520,000 annotated frames and over 18,000 gesture sequences utilising both a custom-collected real-scene dataset and a public gesture benchmark. The model outperforms the existing convolutional-recurrent and lightweight baseline techniques with a macro-averaged F1 score of 95.2% and recognition accuracy of 95.8%. It is a real-time architecture with a comparatively low inference latency of less than 120 ms per gesture sequence. Under a variety of light and backdrop complexity conditions, as well as in situations with ambiguous motions, it can still function effectively and differentiate between known and unknown gesture types. Based on the research findings that VGG16-GRU is both useful and efficient for real-time interactive scenarios, this work offers comprehensive support for modern gesture-driven interfaces.