Algorithm Of Pedestrian Pose Recognition And Behavior Prediction Based On Deep Learning

Pedestrian pose recognition and action discrimination are innovative machine vision tasks and have become a hot research object in many fields.They have good applications in many fields such as intelligent video surveillance,intelligent transportation,human-computer interaction and motion analysis Prospects.More importantly,it represents an over-understanding of the category of image processing from image target detection to image processing,which is a challenging topic.The traditional recognition of human behavior is to obtain the pedestrian mask firstly by the background subtraction,and then the classification task is carried out by template matching or traditional machine learning.However,the background subtraction method is easily affected by the unnatural effects caused by environmental changes.The mask will result in the waste of information.Therefore,the performance of traditional machine learning algorithms has room for improvement in image classification.In recent years,due to the rapid development of deep learning,intelligent surveillance has been able to complete pedestrian detection and statistics,but more detailed posture recognition tasks and behavior determination tasks cannot be achieved.Although there have been deep learning algorithms that can do pedestrian pose analysis,they still remain at the level of doing theoretical research on the single person in the image,and still cannot be practically applied in the face of complex monitoring scenarios.The action discrimination is often based on a single feature for learning judgment,if the feature selection is too much,it will increase the calculation amount,and if the selection is too small,it will cause the accuracy to decrease.Therefore,there are many difficulties in practical applications.In order to solve the above problems,this paper focuses on the multi-person gesture recognition algorithm based on affinity domain,and the behavior discrimination algorithm based on the joint learning of global features and local features.For the study of pedestrian pose estimation,the main research contents of this dissertation are as follows: 1)Using a method of predicting the distribution of heat map of the human body posture,the sample images are sent to the VGG feature extraction network to obtain the features.Obtain the probability heat map of 18 joints including left and right eyes,leftand right ears,nose,neck,left and right shoulders,left and right elbows,left and right wrists,left and right hips,left and right knees,left and right ankles through multi-branch and multi-stage tasks;2)Use an affinity-based approach to solve multiplayer situations in images.The sample image is sent to the Feature Extraction Network shared with the predicted heat map,and the affinity domain of 17 groups of connected nodes,such as the left eye to the left ear,the left eye to the nose and the like,is obtained through multi-branch and multi-stage tasks.3)The acquired heat map of the joint point and the affinity domain of the connected joint point are related by the graph theory-related algorithm,so as to effectively solve the problem of multi-person gesture recognition.For the study of pedestrian movement discrimination,the main research contents of this dissertation are as follows: 1)A novel network based on improved gesture recognition is proposed,including gesture recognition and mask prediction;2)Detection is performed using the full convolutional network branch contained in the new network.Pedestrian mask in the image,which is cut as an input sample and sent to the residual network to obtain the output vector;3)Use the posture recognition branch contained in the new network to predict the 18 joint points of the human body,and then use the neck as the origin to normalize the coordinates An 18-dimensional feature vector is formed and sent to a multi-layer perceptron to obtain an output vector;4)The eigenvectors output from the first two steps are added,and the cross-entropy cost function is used to jointly learn and predict the classification.Finally,the experimental results show that the proposed pedestrian attitude recognition and motion discrimination algorithms based on depth learning all have good performance and can be applied to the actual scene such as video surveillance.