Fall Detection Based On Multi-feature Fusion And Machine Learning

With the increasing aging of population and the development of intelligent pension system,the automatic detection of video based tumble for middle-aged,elderly and disabled people has important theoretical and social significance.The following problems exist in the existing methods of fall detection.Firstly,it is difficult to effectively extract posture features to distinguish static gesture and dynamic movement from falling.Most algorithms still focus on extracting static features or dynamic features unilaterally.It is difficult to distinguish between lying and falling from a static point of view.From dynamic perspective,it is difficult to distinguish between quick squat and fall.Secondly,we cannot guarantee the high accuracy and real-time performance of fall detection at the same time.Identifying the characteristics of fall is determined by multiple parameters.Using geometric statistics or linear discriminant classifier,the accuracy is not high enough.Deep learning can accurately detect fall posture,but because of the complexity of the model,the real-time performance is not guaranteed.Thirdly,it is difficult to identify of fall when the pedestrians are in small scales.Most algorithms focus on detecting and falling down at a single scale.Under the monocular camera,when the source domain is in a large scale,the target area is in a small scale,it is prone to misjudge.In view of the problem of distinguishing the similar attitude and behavior of fall,this paper uses the Open Pose method firstly to take the VGG pre trained network as the skeleton,and two branches return the correlation vector field between the joint point position and the joint point respectively.Then remove the noise according to the center map.In successive stages,each branch is iterated to refine the whole prediction step.A loss is calculated at each stage.Then the joint point position,the correlation vector field and the original input are fused as input for the training of the next stage.After getting the posture feature,this paper presents a new feature of falling down,calculating the static and dynamic motion features respectively,using mathematical statistics to fuse static and dynamic features,and to determine the fall behavior.In view of the accuracy and real time of the fall recognition,this paper proposes a method to further solve the classification problem of complex feature data using two classifiers.From the microscopic perspective,the conventional classifier only set a threshold in the decision level,so as to distinguish the confidence level of each sample.In this paper,we set two thresholds in the decision level of the SVM algorithm,and the samples between the two thresholds are classified into CNN for fine classification.The advantage of SVM algorithm is simple and portable.The advantage of CNN algorithm is high accuracy.Because most of the identification problems can be identified in the SVM link to the positive classification results,only a small number of similarity high recognition problems need to be transferred to the CNN algorithm.Therefore,the two classifier method has the advantage of SVM time efficiency and CNN's high accuracy.Aiming at the problem of multi-scale fall detection,a three-dimensional(could be extended to high-dimensional)scale enhancement model is proposed in this paper.Small scale target's fall feature can be enhanced by the proposed model to avoid miscarriage.Because the feature dimension of the multi feature fusion in this paper is three-dimensional,the three-dimensional coordinate system is constructed.Each dimension is set up a standard body joint point scale,equal interval subdivision scale coordinate,and takes the sub-standard scale.When we recognize the fall,the algorithm normalize the actual measured scale according to the dichotomy to the similar scale,and then scale the scale according to the adjacent rule.The whole step is equivalent to the two interpolation calculation,and then the data is unified into the two classifiers to classify.The above methods have been comprehensively tested in this paper.Some of the results have been published in SCI journal papers.Experimental results show that the proposed method not only has high accuracy in fall recognition,but also has good real-time performance.The new fall feature proposed in this paper can distinguish high similarity behavior from static or dynamic perspective.In addition,for small scale targets that far away from the single camera,the proposed method can adjust the size adaptively to maintain the correct rate of the fall behavior recognition.In conclusion,the method proposed in this paper is superior to the existing fall detection methods in many aspects.