YOLO-AB: A Fusion Algorithm for the Elders’ Falling and Smoking Behavior Detection Based on Improved YOLOv8

The behavior safety testing of more and more elderly people living alone has become a hot research topic along with the arrival of an aging society. A YOLO-Abnormal Behaviour (YOLO-AB) algorithm for fusion detection of falling and smoking behaviors of elderly people living alone has been proposed in this paper, which can fully utilize the potential of the YOLOv8 algorithm on object detection and deeply explore the characteristics of different types of behaviors among the elderly, to solve the problems of single detection type, low fusion detection accuracy, and high missed detection rate. Firstly, datasets of different types of elderly behavior images such as falling, smoking, and standing are constructed for performance validation of subsequent algorithms. Secondly, the Content-Aware Reassembly of Features Module (CARAFE) is introduced into the YOLOv8 algorithm to enhance content perception, strengthen feature fusion, generate adaptive kernels dynamically, and reduce parameters effectively. Then, the Large Selective Kernel Network (LSKNet) module is added to the backbone network part to strengthen the framing of human targets and improve detection accuracy. Next, the Focal-SCYLLA-IOU (F-SIOU) loss function is used to improve the positioning accuracy of the edge part of the target detection frame. Finally, YOLO-AB and other different algorithms are tested and compared using the falling dataset, the smoking dataset, and the falling and smoking mixed dataset, respectively. The results show that the detection accuracy of the YOLO-AB algorithm is 0.93 on the falling dataset alone, 0.864 on the smoking dataset alone, and 0.923 on the falling and smoking mixed dataset, all of which are better than those of the other algorithms. The performance of YOLO-AB is better than that of YOLOv8 on multiple metrics, such as 4.1% improvement in the mAP50 index, 4.9% increase in the P index, and 3.5% boost in the R index, which verifies the effectiveness of the algorithm.