| With the rapid development of China,the aging of Chinese society has become a new social problem.In order to help the elderly to obtain corresponding help in time when dangerous behaviors such as falling,going up and down stairs,running or abnormal actions occur,and avoid serious consequences.Due to the behavioral characteristics of accidental,sudden,and uncertain behaviors of the elderly;secondly,the daily pictures of the elderly’s behaviors and actions taken by cameras will also be affected by factors such as light,shadows,and occlusion;in addition,the behavior of the elderly It not only has the characteristics of space dimension,but also has the characteristics of time dimension.In summary,there are these difficulties.This article aims to help the elderly who cannot get help in time for the occurrence or face of risky behaviours by developing the computer field which is based on the deep learning area,then applying to the relevant knowledge of human pose estimation,and finally proposed the three-dimensional convolution neural network based risk behaviour recognition method for the elderly.This paper mainly introduces the following three aspects in accordance with the order of advancement of the research:(1)network Lenet-5 based on the classic image classification.(2)Image classification network suitable for transfer learning based on Vgg Net.(3)Based on convolution neural network(CNN)combined with long-short term memory model(long-short term memory),referred to as LSTM,for research.this paper is creative,the first technology of human pose estimation is used to convert a segment of human behaviour into human behaviour and pose map,then use 25 continuous human behaviour and pose maps as the input of the behaviour recognition;Due to the normal image classification model used the method of two-dimensional convolution,and the data dimensions during input are inconsistent,in this case,the calculation method of three-dimensional convolution is introduced.After many experiments,several different three-dimensional convolution neural network models were constructed.For now,only a continuous action sequence of the elder human behaviour is needed to continuously predict the current behavioural state.By comparing data,analyzing results,evaluating models,and multiple experiments,the deep learning-based method for identifying dangerous behaviours of the elderly can better identify behaviours of the elder people.test accuracy of several different three-dimensional convolution neural network models were all above 87%.Several risky behavior identification methods for the elderly proposed in this paper can effectively identify and detect dangerous behaviors such as falling,going up and down stairs,and taking medicine for many times,as well as other normal behaviors.When the elderly encounter dangerous behaviors,the first Send relevant help information to family members and related personnel in a timely manner to ensure the life,health and safety of the elderly in a timely manner. |
