Visual Detection Algorithm For Personnel Abnormal Operation Based On Deep Learning

Abnormal operation detection and early warning is one of the important indicators to measure the safety production.The traditional anomaly detection algorithm uses the manual feature or the shallow neural network to evaluate the safety level of the operation.Due to factors such as workshop conditions,light difference and model capabilities,traditional algorithms often fail to achieve the desired accuracy and robustness performance.In this paper,a visual detection algorithm for personnel abnormal operation based on deep learning is proposed,which includes two parts: personnel detection and behavior identification.Firstly,a traditional personnel detection algorithm based on spatial pyramid and integrated features is proposed,by analyzing its accuracy and real-time performance,a lightweight detection network based on feature fusion is proposed.In view of complicated recognition problems of behavior sequences,high precision recognition is accomplished by combining deep residual network,3D convolution operation and binary direction long short term memory neural network.Finally,with reasonable allocation of computing resources,a load-balancing framework based on functional separation is established.The main work of the paper is as follow:(1)A traditional personnel detection algorithm based on spatial pyramid and integrated features is proposed.To address the problem caused by different sights,the algorithm introduced multi-scale detection by constructing Gaussian spatial pyramid space through the original image.Furthermore,a dual-channel serial detection framework was proposed,preliminary detection utilized HOG-PCA basic detector and linear support vector machine which had fast speed and low undetected rate,an advanced RGB-SIFT-PCA detector and random forests with high precision and low error rate were used for secondary screening.Finally,non-maximum suppression algorithm was adopted to remove redundant bounding boxes.(2)Improve the above algorithm,a lightweight detection algorithm based on feature fusion is proposed.Aim at poor real-time performance of traditional neural network in low end hardware platform,convolution layers reduction and Xception structure are proposed to compress network parameters.Aim at the low feature richness caused by parameters reduction,introduce mid-high level feature maps fusion through shortcut connection and dimension reduction,and finally improve detection performance under the condition of different scale.(3)An abnormal operation recognition algorithm based on deep spatio-temporal network is proposed.This algorithm aims to improving the accuracy and reducing computing cost on the basis of tradition model,such as two stream convolution network and convolution 3D network.Parameters-shared deep residual network is proposed as a pre-feature extractor to extract spatial features of image frames.A method with 3D convolution and binary direction long short term memory network is proposed to extract temporal features to accomplish timing operations recognition.(4)Apply deep learning technology to workshop personnel abnormal operation detection.Analyse actual hardware platform,a load-balancing framework based on functional separation is proposed.The lightweight personnel detection network is placed in the robot client with low end configuration to gain real-time performance and computing parallelism.The abnormal detection network is placed in high-performance server to achieve high precision analysis of time operations.Through creating multi-running instances in server to boost system's parallel ability and reduce response time.