Research On Human Fall Detection Method Based On Convolutional Neural Network

Human fall detection is an effective means of preventing death and secondary injuries caused by falls in the elderly,and is a powerful safeguard for people’s health in their old age.Most of the current fall detection technologies and devices are not mature,and most of the computer vision-based detection technologies are generally not high in real-time and accuracy.At the same time,due to the complexity of the application scenario,the diversity of human behaviour and the cost of hardware,it is still very challenging to detect falls quickly and provide timely feedback.In recent years,due to the rapid development of deep learning technology,researchers at home and abroad have started to try to apply it to the field of human behaviour recognition to improve the accuracy of detection and the adaptability of algorithms.In this paper,we use lightweight convolutional neural networks as the basis for target detection and key point recognition of human body,and carry out various fall detection methods and data fusion,taking corresponding theoretical and experimental research,the main research contents are as follows:(1)Compare the characteristics of traditional fall detection systems and visionbased fall detection,and construct theoretical ideas for fall detection.Starting from the advantages and disadvantages of traditional fall recognition methods,a computer vision approach is used to establish a fall detection scheme suitable for indoor use.Different convolutional neural networks are learned and studied,and the feasibility of lightweight convolutional neural networks for human key point detection in mobile is verified,and a human fall detection method based on spatio-temporal fusion with multiple criteria is proposed based on lightweight human key point recognition network.(2)A temporal model corresponding to the spatial fall behavior detection model is established to classify the fall behavior.The completion of the fall action is not only a continuous process of spatial change of the relative position of the human body,but also corresponds to continuous temporal change.In this paper,the two are linked by means of interval video frames to solve the correspondence between temporal and spatial changes before and after the fall.From the two perspectives of the relative position of human body and camera and the orientation of human body falling,the sitting and standing falling behaviors are classified,and the common features are analyzed and summarized for different falling behaviors.(3)Identify the human key points and establish a spatial fall recognition model.A lightweight network structure is used to identify the key points of the human body.The key points are tracked in real time and the coordinate data are visualized,and the key points are selected to form an appropriate vector to represent the human body.The waveforms generated by the falling behavior are observed and analyzed,and compared with common walking and sitting behaviors.The mathematical calculation model is built from the changes in various aspects of data such as angle,mode length and height-towaist ratio caused by human falling behavior.The temporal and spatial model parameters of fall recognition are parameterized using UR and Le2 i datasets,and the complete feature extraction and fall recognition model is obtained by adding coordinate complementation and window function.(4)The recognition model is ported to the built experimental platform for dataset experiments with actual complex scenes.The experimental results show that the recognition effect is good in two different datasets,which can reach 92%.Under the same lighting conditions,multiple sets of experiments were conducted in different types of complex rooms,and the recognition rates of standing and sitting falls could reach 95%and 85% on average,respectively,and the experiments showed that the model has high adaptability in different complex environments.Figure [45] table [6] reference [80]...