Research On Fall Detection Based On Machine Vision

As China’s population ages and the phenomenon of "empty nesters" becomes more common,more attention is being paid to the health of the elderly.Accidental falls are a major cause of injury among the elderly in their daily lives.Fall detection for the elderly is becoming a popular area of research and the rapid growth of the elderly population is providing a large demand and market for such products and technologies.Fall detection solutions based on machine vision have become a hot topic of research in fall detection due to their ease of installation and wide application.This thesis provides an in-depth study of the algorithms involved in the process of machine vision-based fall detection for the elderly.The main research includes the following aspects.(1)A fall detection dataset is produced.The datasets used in this thesis include the publicly available dataset Le2 i and the UR Fall Detection Dataset,in addition to some home-made datasets.These datasets are uneven in terms of scenarios and number of scenarios.To address this issue,the datasets were randomly augmented and different scenarios were combined to form a hybrid dataset.Accuracy rate and recall rate were selected as the evaluation metrics for human target detection,and accuracy rate and fall detection error rate were selected as the evaluation metrics for fall detection models.(2)A fall detection algorithm based on traditional machine learning was investigated.A fall detection scheme combining HOG(Histogram of Oriented Gradients)features and support vector machine algorithms is proposed.For human target detection,HOG combined with SVM(Support Vector Machines)classifier is used for target detection.For fall detection,a two-stage SVM classifier is proposed for fall detection.The first stage SVM classifier eliminates normal behaviours in daily life,and behaviours that cannot be identified by the first stage SVM are input to the second stage SVM for fall detection to identify fall behaviours.The first stage SVM classifier uses a linear kernel with higher accuracy,and the second stage SVM classifier uses a RBF kernel with higher classifier accuracy.The SVM classifier with RBF kernel function is parameter-seeking optimized using the particle swarm optimization algorithm,and its fall detection accuracy reaches 85.8%,Improved accuracy by 4.5%after optimization compared to before optimization.(3)A fall detection algorithm based on deep learning was investigated.By comparing the current single-stage and two-stage target detection algorithms,the use of the YOLO target detection algorithm was identified.Furthermore,the YOLO network model was improved using the Dense Net algorithm idea.Experiments demonstrated that the improved model had higher detection accuracy and no significant increase in parameters.In terms of human pose estimation,human key point extraction was investigated.The Alphapose algorithm is optimised for the problem of incomplete detection of human skeleton joints due to the presence of occlusion in the target,and the key point complementation algorithm with geometric and temporal constraints is proposed to optimise the Alphapose algorithm,acquire 0.3% improvement in detection accuracy.The fall detection algorithm was studied using spatio-temporal convolutional maps,and the Kalman filtering algorithm was used to optimise the algorithm for the problem of fluctuating joint point trajectories caused by environmental disturbances,the joint point trajectory accuracy was improved by 2.2%,The final fall detection experiment accuracy reached 91.3%.(4)A fall detection system was designed using Qt5.The fall detection system includes functional interfaces such as login,fall detection,result export and fall alarm,and comparative experiments were conducted between the HOG+SVM fall detection algorithm and the deep learning fall detection algorithm under the influence of different lighting conditions and occlusion.The experiments show that the two fall detection solutions proposed in this thesis can achieve high detection accuracy and can meet the needs of elderly fall detection.The deep learning-based fall detection scheme is less affected by illumination and occlusion,and is overall better than the traditional machine learning algorithm fall detection scheme.