Research On Sitting Posture Visual Recognition Method Based On Discriminant Deep Learning

Good sitting posture has an important impact on improving people's life,work efficiency,maintaining their physical and mental health.Because of its advantages in image feature extraction and classification,deep learning is applied to the method of sitting posture visual recognition in this paper.Firstly,this paper makes some improvements on the activation function and loss function in convolutional neural networks,and then applies them to the recognition of sitting posture images via monocular vision and binocular vision,respectively.The main work of this paper is summarized as follows:1.This paper makes some improvements on activation function of traditional convolutional neural networks to address some problems of the existing activation function.This paper proposes a new Softrelu activation function by combining the advantages of the LReLU function and the Softplus function to improve network performance.Moreover,since the learned features via classical CNNs may have poor discriminant capability for image recognition,this paper introduces a new linear discriminant analysis loss(LDloss)into CNNs to explicitly achieve intra-class compactness and inter-class separability among learned features.Experimental results on different data sets verify the effectiveness of the two proposed improved algorithms.2.The paper applies these proposed methods to the monocular sitting posture recognition.Firstly,the face detection method is adopted to provide the calibration input of the automatic segmentation algorithm to achieve the sitting posture target foreground extraction.Then,this paper adopts the improved CNN to learn the deep features for sitting posture recognition.Different experiments have verified the efficient on monitoring and distinguishing people's sitting positions of the proposed algorithm.3.The paper further applies these proposed methods to study the binocular sitting posture recognition since the binocular camera can effectively provide the depth information of images.This paper firstly adopts binocular cameras to obtain the disparity map,and then combines the disparity map with the left and right sittingposture images collected by the binocular camera as the inputs of improved CNN.Finally,the improved CNN is applied to the feature extraction and classification recognition of sitting images.The effectiveness of the proposed algorithm is verified in the binocular sitting recognition data set and the real-time posture recognition video,respectively.