Elderly Fall Detection Based on Improved SSD Algorithm

Abstract

We propose an improved a single-shot detector (SSD) algorithm to detect falls of the elderly. The VGG16 network part of the SSD network is replaced with the MobilenetV2 network. At the same time, we change the infrastructure of MobilenetV2 network, the three layers that were not downsampled at the end were removed, which can make the model structure simpler and faster to detect. The complete Intersection-over-Union (CIoU) loss function is introduced to get a good regression of the target borders that have different sizes and different proportions. We use Feature Pyramid Network (FPN) for upsampling, it can fuse low-level feature maps with high resolution and high-level feature maps with rich semantic information. For sampling results, we use the Secure Shell (SSH) module to extract different receptive fields, which improves the detection accuracy. Our model ensures that the accuracy of the elderly fall detection remains unchanged, but it greatly improves the detection speed that only takes 10 milliseconds to detect a picture.