Research On Home-based Fall Behavior Detection Based On Improved Yolo V5s

China has entered an aging society since 2021,and with the current trend,the aging problem of the society will be further aggravated.The fall problem of the elderly is one of the important factors affecting the health of the elderly.The fall events of the elderly living alone occur frequently.If the elderly can’t get timely help after falling,it may be injured or delayed treatment,or it may endanger life in serious cases.Therefore,the research with real-time fall detection is very necessary.In this paper,the characteristics of human standing and falling posture are extracted,and the attention mechanism is introduced to focus on the body information.The main contents of the study are as follows:(1)The human target recognition algorithm based on fusion channel spatial attention mechanism was studied.In the monitored environment,it is necessary to first obtain the position of the human body in order to make accurate judgments about the posture of the human body.This paper proposes a fusion channel spatial attention mechanism to address the problem of difficult recognition of human bodies due to complex indoor environments and a large number of objects.The channel relationship dependency learning layer of CBAM utilizes 1 × The convolutional layer replaces the original fully connected layer,reducing the number of parameters in the model.After adding attention mechanism to the network residual block structure and backbone network,it can pay more attention to the movement of the human body in the area,and effectively reduce the interference of other objects to the results.(2)Regarding the existing Yolo v5 s network,several modifications have been made.For the backbone network,the Leaky Re LU activation function has been replaced with the hard-swish activation function.The original SPP structure has been replaced with the ASPP network structure.In terms of the neck structure,the Bi FPN structure has been adopted,along with the pre-Bi FPN structure for further feature extraction,replacing the FPN+PAN structure of Yolo v5 s.Through comparative experiments,these superior modifications have significantly improved the precision and recall rate of object detection.(3)A fall detection method based on video frame timing change is studied.The general fall detection based on video is often to cut the image of each frame to classify its posture.After the study of these algorithms,it is found that such algorithms are easy to cause misjudgment,such as prone to consider lying flat,squatting and other postures as fall states.This article combines object detection with multi frame correlated video frames,designs an algorithm for fall detection,and adopts a multi camera voting mechanism for comprehensive judgment,effectively reducing the false positive rate.