Recognition Of Fall At Home Based On Scene Analysis And Activity Classification Using Deep Learning

Automatic human fall detection is an important research topic in caring for vulnerable people such as elders at home and patients in medical places.Over the past decade,numerous methods aiming at solving the problem were proposed.However,these existing methods based on computer vision have these following defects.(1)These existing methods only focus on detecting human themselves and lack of the analysis about scene.Therefore,the existing methods are difficult to distinguish between some fall-like activities and special fall activities.These methods are also easy to produce false detection and error inspections.(2)Most existing methods are proposed to class falling and non-falling problems so that the falling behavior cannot be identified furtherly.(3)These existing models based on convolutional neural network are complex,redundant and hard to train,so that these methods have defects in spatial storage and detection efficiency.As for lacking scene analysis in existing methods,an algorithm of feature extraction based on scene analysis is proposed in this paper.The method first uses object detection framework Faster R-CNN to detect scene objects at home,including human,sofas,chairs and so on.To alleviate the situation of missing detection and occlusion,a scene prediction algorithm is proposed based on object position detected in previous frame to predict missing objects.Meanwhile,the activity characteristics of detected people such as human shape aspect ratio,centroid,motion speed are detected and tracked.As for no further identification of falls in existing methods,an automatic decision engine algorithm for fall recognition based on Gaussian Mixture model is proposed in this section.By fitting each behavioral feature to different Gaussian distribution,the methods can effectively determine the behavior clustering center of each fall behavior.Thus,many falls and other daily behaviors can be identified accurately.For the traditional CNN model,there are problems in complex models,large parameters and difficult training,a two-stream channel fall classification model based on motion characteristics and mobile VGG network is proposed in this paper.One channel stream is the motion feature of human for predicting falls.The other channel stream uses progressive mobile VGG to class falls and daily behaviors.The model combines motion characteristics to design a two-stream lightweight mobile VGG for fall classification.At the same time,residual connection between layers is applied to overcome the disappearance of shallow parameter gradient in deep model and the hindrance of gradient reflow.The network is also optimized for pooling layer and full connected layer.Sufficient experiments have been carried out on above proposed methods.Many experiments show that the feature extraction algorithm based on scene analysis can well represent scene features and behavior characteristics.Especially for some special falls and fall-like activities.Combined with scene analysis,our proposed method not only accurately and effectively detects falls on furniture such as sofa and chairs but also distinguish them from other fall-like activities such as sitting or lying down,while the existing methods have difficulties to handle these.In addition,our proposed two-stream fall classification model reduces model storage occupation and improves the efficiency of fall detection.