Research On Gesture Recognition Method Based On Multi-feature Fusion

Contactless gesture recognition method based on computer vision,as a natural and convenient way of interaction,plays an increasingly important role in human-computer interaction.On the other side,the introduction of the depth information in the RGBD images further overcomes the influence of the background complexity and environment illumination changing,which can improve the accuracy of the gesture recognition algorithm.However,how to make full use of the plentiful texture information and depth information in RGBD images is a problem that needs to be considered.In recent years,the emergence of deep learning techniques such as convolutional neural network has provided an effective approach for the use of RGBD images.Feature extraction based on deep learning can help to find more accurate and robust feature representations for RGBD images.Moreover,we can further optimize feature information and improve the performance of gesture recognition algorithms through the complementary fusion of local feature descriptors and deep features.This research is launched based on the above analysis.First of all,a parallel connection structure of multi-channel convolutional neural network is proposed in this paper.The algorithm offers a new way to extract unified-paradigm features though the coupling connection and weight balance across RGB and depth channels,which realizes the deep combination of texture information and 3D information.At the same time,the deep-feature learning model is not limited to RGBD format datasets due to the flexibility of its network construction.The model can achieve the deep-feature extraction of various multi-modal datasets through the construction of multi-channel and coupling training.Secondly,a novel multi-feature fusion method is proposed.The algorithm achieves the weight balance among various features through the fusion of locally reconstructed weight matrix in each feature space.At the same time,the algorithm introduction a mapping transformation between the input feature space and output feature space,which can reduce the high spatial and temporal costs in practical applications.Moreover,the proposed method utilizes the label information in the optimization target of multi-feature fusion algorithm,which helps to narrow the intra-class distance and enlarge the inter-class distance.Then,the original problem is transformed into an unconstrained optimization problem and a quadratic programming problem.The experimental results also show that the proposed method based on deep network and multi-feature fusion is reasonable and effective.