The Technology Research Of Gesture Recognition Based On Neural Networks

With the computer in terms of application areas continue to expand, gesture-based interactive behavior that gesture recognition technology has been widespread concern. Gesture recognition systems can generally be divided into the following areas: image preprocessing gestures, gestures segmentation, feature extraction and hand shape recognition, gesture tracking, trajectory feature extraction and recognition. The research work in this thesis is based on neural networks, focusing on the depth study of the various modules gesture recognition system.The main work in this thesis are as follows:1. For the problem of the current de-noising algorithm can only gesture to remove a noise, or the removal of a variety of noise, they also damage the integrity of the gesture edge features, this thesis presents a new de-noising algorithm which based on pulse coupled neural networks and complex de-noising algorithm. Experiments show that the algorithm can effectively remove a variety of noise, while preserving the integrity of the gesture edge features.2. In the filed of gesture segmentation module, the cellular neural networks is used to split gesture. Experiments show that this method has better segmentation effect gesture segmentation. Due to the geesture segmentation based on cellular neural networks is easily to be influenced by other gesture segmentation edge from the background information, so this thesis proposed a method for network segmentation gesture based on the cellular neural networks and hand shape features. Experiments show that the method can effectively remove the interference of background information, but also has a good gesture segmentation.3. In the filed of hand feature extraction and recognition,we improve the method of the traditional Fourier descriptor, the improved method retains the phase information of Fourier descriptors, which makes this method different rotation angles produce different gestures feature vector. Then, we use the BP neural network to identify the hand operation. Experimental results show that the improved Fourier descriptors feature extraction methods can solve the problem of the rotation gesture, and the gesture recognition algorithm used BP neural network algorithm also has a good performance.4. In the filed of gesture tracking, we improve the method of the traditional particle filter tracking algorithm for solving the target occlusion and background color interference problems. First, we design a particle filter algorithm which based on edge characteristics as the target model to eliminate the interference of background color; then, the method and Kalman filter combined with the predicted Kalman filter tracking mechanism to solve the occlusion problem. Experiments show that the improved particle filter tracking algorithm can solve the occlusion problem gesture, and also has a better tracking performance.5. In the filed of gesture trajectory extraction and recognition, we use the orientation angle and position of the features characterizing the trajectory feature; then, the Freeman chain describes the discrete data of gesture trajectory; and finally trajectory recognition operations with hidden Markov models. Experimental results demonstrate that our system can recognize dynamic gesture trajectory in real-time and has a good performance.