Research On Fall Detection Algorithm Based On Deep Learning

Nowadays,Falls occur frequently in daily life and there are numerous injuries and safety hazards caused by falls,making fall detection a study of great importance.If a timely response can be made when a fall event occurs,the injuries caused by falls can be reduced.Therefore,this thesis proposes a fall detection algorithm based on deep learning,and the main research content is described below.(1)Since the characteristics of fall actions and lying down actions are extremely similar,the current public dataset lacks data on lying down behaviors.This study establishes a human action dataset from the surveillance viewpoint,which mainly contains fall,lie down,squat,and sit down,and establishes the image data of limb occlusion.This thesis adopts a fall determination strategy combining spatial dimension and temporal dimension.By extracting the human body aspect ratio,human body tilt angle,head centroid velocity,and center of mass velocity,a multi-feature fusion strategy is proposed to determine the falling behavior,with shape features in the spatial dimension used to distinguish daily behaviors such as falling,squatting and sitting down,and motion features in the temporal dimension used to distinguish falling and lying down.A large number of experimental analyses are conducted on the self-built dataset,which proves that the proposed fall features can effectively distinguish between falls and daily behaviors.(2)For the case of limb occlusion during motion,the images with limb occlusion are discussed and a symmetry-based human keypoint complementation method is proposed.By using the key point location information output from OpenPose,the missing human key points are ranked,and the symmetric key points are used to derive the location information of the missing key points and to fill in the missing key points with the human body centerline as the symmetry axis.We propose an evaluation index ADV for the algorithm,and calculate the distance difference variance between the real key point coordinates and the alignment point coordinates to evaluate the algorithm.(3)In this thesis,we construct a multi-feature parameter fall detection method,firstly,we perform target detection on video sequences,improve YOLOv7,improve the E-ELAN layer of the backbone extraction network in YOLOv7 by adding the SE module,and use SIoU to replace the original border regression loss function,the mAP@0.5 of the improved model increased by 1.45 percentage points,the recall rate Recall improved by 1.11 percentage points;OpenPose is used to estimate the pose of the human body and obtain the information of fall features,and the sequence information of video is processed using a classifier to effectively distinguish fall behavior from daily behavior.The experimental results prove that the algorithm can reach 94.8% accuracy on the self-built dataset and 95.6% accuracy on the public dataset Le2i,which can meet the basic requirements of fall detection.