Research Of Falling Detection Method Of Pedestrians Taking The Escalator Based On Human Pose Recognition

With the rapid development of the national economy,the pace of urbanization is accelerating,and the number of escalators in use is rapidly increasing.In places where large passengers,subway stations,railway stations,and airports are concentrated,the number of passengers is placed.A considerable escalator.However,with the rapid increase in the number of escalator equipment,the frequent occurrence of accidents related to escalators has caused great damage to the personal safety of pedestrians.In order to solve the above problems,a device capable of accurately detecting the fall of a pedestrian on an escalator is indispensable.Since most of the fall research at this stage is the use of wearable devices or PC-based recorded video,lack of flexibility and real-time,this paper proposes a real-time fall detection method based on computer vision,the Open Pose model and feature pyramid The network model(FPN)is fused by predicting the complete human body by detecting the joint points of the human body and combining the partial affinity domains,and then combining the human body parameters to establish feature points and calculating the threshold value to determine whether the human body falls.Through this method,a real-time and feasible fall behavior recognition algorithm can be realized.The main work of this paper includes the following aspects:(1)Based on the characteristics of the human body,the judgment conditions are set for the fall of the pedestrian when riding the ladder and other common postures.Firstly,it analyzes people's fall behavior on escalators and several other common behaviors.It classifies the state of pedestrians when they are on the ladder,and combines the classification results with the causes and processes of pedestrians falling on the ladder to summarize the performance of different states.Feature,the feature is combined with the human body's own parameters,and then seven feature points can be established according to different situations of the actual scene.By obtaining the feature point coordinates to obtain the correlation angle value and the vector mode length,the degree and modulus length of the angle are matched.Different threshold conditions are set to judge the pedestrian's erection,bending,walking and falling behavior.(2)A fall detection algorithm based on the FPN?Open Pose model is proposed.Based on the Open Pose model,this model proposes a solution to integrate the feature pyramid network(FPN)with Open Pose for Open Pose's existing problem of poor detection of small targets and abnormal poses,and strengthens the detection effect on small targets.And through the Tensorflow deep learning framework,the algorithm is trained in the way of migration learning,using a manually labeled data set containing abnormal human behavior,and then the output of the human joint point coordinates is judged by the above mentioned method.Finally,the fall detection algorithm based on the FPN?Open Pose model is obtained.(3)Experiment and analyze the pedestrian fall detection algorithm based on FPN?Open Pose.First,build the environment required by the algorithm,and then use the new manually labeled data set to train the FPN?Open Pose model proposed above through migration learning.Next,according to the characteristics of the actual scene,four different conditions of the pedestrian fall detection algorithm based on FPN?Open Pose are set to detect the four states of pedestrians when riding,and the effect of the Pedestrian fall detection algorithm based on Open Pose before the improvement is improved.Compared with the algorithm proposed in this paper,we can get the conclusion that the proposed algorithm is better.