On improvement of fall detection algorithm based on YOLOv8:CASL-YOLO
Falls were considered as one of the common accidents with dangerous harm to elderly people,and also were primary cause of disability and injury death for the elderly people over 65 years old in China.How-ever,the current mainstream fall detection technology was greatly affected by the environment,the detection accuracy was low in complex scenes such as object occlusion and illumination change,and the model parame-ter quantity and calculation amount were high,resulted in high cost.It then was not well deployed and applied to real life scenarios.In order to solve the above problems,it was proposed a lightweight fall detection algo-rithm based on YOLOv8 model in complex environments:CASL-YOLO.Firstly,the spatial depth convolution(SPD-Conv)block was introduced into the model to replace the traditional convolution module,which re-tained all the information in the channel dimension by convolution operation for each feature map,so as to im-prove the performance of the model in low-resolution images and small object detection;Secondly,it was intro-duced an attention mechanism based on location information to capture cross-channel,directional and location-aware information,so as to locate and recognize human targets more accurately;Finally,the selective large convolution kernel(LSKNet)was introduced into the feature extraction module to dynamically adjust the receptive field to effectively deal with the complex environmental information in the fall detection scene and im-prove the perception ability and detection accuracy of the network.Experimental results showed that on the public Human Fall dataset,CASL-YOLO's mAP@0.5 reaches 96.8%,which was better than the baseline YOLOv8n.At the same time,the model only had 3.4 MiB parameters and 11.7×109 computation.Compared with other detection algorithms,CASL-YOLO achieved higher accuracy and performance with a small increase in the number of parameters and calculation,and meets the deployment requirements of actual scenarios.
fall detectionYOLOv8attention mechanismspatial depth convolutionlarge selective kernel network
NETL
NSTL
万方数据
