Speed Invariant Hand Gesture Recognition System using CNN

Feature extraction process in dynamic hand gesture recognition systems sometimes become unsuccessful due to the complexity of the dynamic hand gestures. Speed variation is one of the challenges faced by the existing dynamic hand gesture recognition systems. In this paper, different deep learning networks have been used to recognize gestures which lead to avoid feature extraction process. Hence, complexity and computation time is also decreased for the recognition of the dynamic hand gestures. Comparison of various deep learning algorithms has been done for the dataset consisting of numerals from 0 to 9. Hand detection followed by the tracking of gestures is done by two-level KLT tracker and gesture trajectory is being used as input to the deep learning algorithms. It has been observed that 99.33% recognition accuracy is achieved using Resnet18 and InceptionV3 deep learning architectures. Also, least computational time is obtained using Vgg16 algorithm.