Research On Target Detection Algorithm Based On Human Pose Information

Target detection algorithm has always been a favorite in the field of artificial intelligence,because target detection can be used in industry,medical treatment,entertainment or smart home.In the past few decades,target detection technology has developed from traditional manual feature extraction to detection using deep learning algorithm.The algorithm has evolved step by step in the direction of more and more intelligent model.Most of the existing target detection methods are simple detection methods.Regardless of the influence of the scene,they only pay attention to the target itself,and train the network to directly identify and locate the target.These algorithms have a wide range of application scenarios,but lack of pertinence.Therefore,in the case of the generalized scene oriented by the existing detection methods,this paper selects the scene with action interaction between people and objects as the specific detection background,and proposes a target object detection method based on human posture information.When there is more accurate gesture information of the human body and the target to be measured in the scene at the same time.Firstly,aiming at the problem of how the human posture information is introduced into the model,this paper preprocesses the dataset used,and marks the joint points of the people who interact with the target in the image one by one.The body joints heatmap of human joints with16 joint points is generated by Gaussian heatmap,and the body joints heatmap is added to Center Net model as a branch.Due to the introduction of human posture information,the prior information provided by the joint point thermal map will also be extracted in the process of feature extraction and fusion,so as to improve the detection accuracy.After ablation,the accuracy of joint detection is slightly improved compared with that of the original network,but it is still lower than that of the original network.Secondly,aiming at the problem that the model training becomes slower and the real-time performance becomes worse after the branch of body joints heatmap is added,the Hourglass-104 is lightweight improved to reduce the network parameters and avoid sacrificing speed for accuracy.Experiments show that the computational efficiency of the improved model has been greatly improved,and the detection speed exceeds the original Center Net model of 3.17 FPS.Finally,in order to further improve the feature capturing ability of the network and enhance the recognition accuracy of small targets,the Global Context Block is introduced into each Hourglass network and between the two model.In order to balance the detection efficiency and complexity of the network and avoid adding network parameters again after lightweight improvement,a lightweight Global Context Block is specially selected.Experiments show that the model added with the global context module not only has the same speed as before,but also improves the detection accuracy.Through the overall improvement of the model,a model with better detection effect in special scenarios is obtained.Compared with the initial model,the speed of the improved model is increased by 6.8% and 4.78 FPS,which finally proves the effectiveness of the proposed model.If the model is applied to the detection of hand-held dangerous objects in airports,industrial exclusion zones and other scenarios,it will have important practical significance.