A Method Of Attitude Estimation For A Specific Character

With the development of artificial intelligence,today's artificial intelligence devices have more and more demands for human-computer interaction.In addition to the commonly used language interaction and keystroke interaction,human body movements also contain a large amount of information,such as traffic police directing traffic,service robots analyzing customer behavior needs,coaches analyzing athletes'movements and so on.Therefore,it is necessary to design the appropriate human posture estimation algorithm to enable the machine to understand the movement information of a particular person.In recent years,with the upsurge of deep learning,posture estimation has become a hot topic for researchers.With the deep discussion of the researchers,the human attitude estimation algorithm is gradually becoming mature.When it comes to solving the key problems of posture estimation,such as crowd crowding,human body shielding and computer power limitation,the solutions proposed by the researchers are more ingenious and have their own advantages.However,most good attitude estimation algorithms need to be calculated in parallel in a high-configuration machine environment,and the low-configuration machine environment,which is more common in People's Daily life,cannot provide sufficient computational support for these algorithms.Therefore,it is of great significance to design an attitude estimation system which can be used on a lightweight device and has high practical value.Aiming at the above problems,this paper proposes a method to estimate the posture of the body of the specified person,and designs a set of simple and efficient algorithm to estimate the posture of the body of the lightweight equipment.For an input image,first detect the face through MTCNN algorithm,extract the face frame and face key points;Then,Facenet is used for face recognition to find the objects to be observed.Based on the proportion relationship between human face and human body,the position of the upper body of the observed object is roughly calculated,and then the figure extraction image required by the attitude estimation model input is cut out;Finally,the cropped image was input into the neural network to predict the key points of the body of the specified person.The key points are connected into the shape of the matchstick for the user to observe,and the human body movements are judged according to the position of the key points,so as to achieve the purpose of attitude estimation.For the structure and training method of neural network,a simple and efficient upper body posture estimation algorithm is designed.Inspired by the target detection algorithm YOLOV3,this calculation has the advantage of small computation,fast computation speed,and is suitable for lightweight devices with lower configuration.The PaddlePaddle deep learning platform was applied in the experiment,and the algorithm designed in this paper and the whole system were simulated in a low-configuration environment.The convergence of algorithm training process is demonstrated and the feasibility of the algorithm is proved.The superiority of the algorithm is proved by comparing it with other algorithms in the aspects of operation consumption time and prediction accuracy.The whole system is simulated with traffic police action estimation as a case,and the accuracy rate reaches 72.6%,which proves the practicability of the system.Figure[20]table[4]reference[23]...