DSC-OpenPose: A Fall Detection Algorithm Based on Posture Estimation Model

The fall events in crowded places are prone to public safety problems, where real-time monitoring and early warning of falls can reduce the safety risks. Aiming at the problems of large scale and poor timeliness of existing fall detection methods based on pose estimation, an OpenPose human fall detection algorithm called DSC-OpenPose is proposed, which incorporates an attention mechanism. Using DenseNet dense connection idea as reference, each layer is directly connected to all previous layers in the channel dimension to achieve feature reuse and reduce the size of model parameters. In order to get the spatial direction dependency and precise location information of the feature map and to increase the pose estimation accuracy, the coordinate attention method is introduced between various stages. The method is proposed to identify fall behavior based on human outer ellipse parameters, head height and lower limb height together to achieve fall detection of human targets. It is showed that the algorithm achieves a good balance between model size and accuracy on the COCO dataset. The fall detection approach simultaneously achieves 98% accuracy and 96.5% precision on the RF dataset, reaching a detection speed of 20.1 frames/s. The model is small enough to support the real-time inference requirements of embedded devices.