A Novel Fall Detection Methodology Based on Feedback-Fusion and Optimal-Allocation Model
Bing Qi, Menghan Guan, Weiqi Sun · 2024
In order to offer timely rescue measures for elderly and reduce secondary hurt caused by fall, a novel fall detection methodology based on Feedback-Fusion and Optimal-Allocation (FF-OA) model is proposed. First, K-Means++ algorithm is shaped to re-cluster the anchor to increase detection precision, because the initial anchor is unsuitable for all fall detection dataset. Second, Feedback-Fusion model is used to add a feedback to the Third Stage in Cross-Stage Partial Network (C3) to form the C3-F model, and its feature information is sent back for further fusion processing after each bottleneck stack, which solves the problem small feature of the dataset is ignored in training to cause that targets cannot be identified precisely. Third, Optimal-Allocation model is used to eliminate the detection errors of fuzzy labels caused by anchors overlap in the process of target detection. Minimum loss function is applied to select the optimal anchor for input image to detect fall correctly. Then, hardware platform is designed and implemented to collect the pictures and videos to test detection performance of the model. After being trained and tested, precision and recall of the novel methodology are increased by 2.88% and 2.38% respectively, and [email protected] is increased by 1.42%. Besides, its detection precision is increased to 83.15% and its confidence is increased above about 85%. This methodology is able to provide timely assistance for the elderly, and effectively reduces the risk of dangerous situations caused by the fall, which obviously improves their safety and quality of life.