Cross-Resolution Feature Fusion for Fast Hand Detection in Intelligent Homecare Systems

Abstract

Fast hand detection plays an important role in Intelligent Homecare systems due to its close association with many human-related tasks, such as human activity and behavior recognition. In the last decade, although hand detection has been studied widely, fast hand detection has remained a challenge. In this paper, we employ a single shot multibox detector (SSD) as the base architecture and propose a novel cross-resolution feature fusion (CFF) approach to add contextual information and semantic information to shallower layers for fast hand detection. Our approach helps improve performance significantly, especially in small instances, owing to the inclusion of two important modules: a narrow atrous spatial pyramid pooling (N-ASPP) module and a richer semantic information generation (RSIG) module. The proposed N-ASPP module employs atrous convolution to capture multiscale context information by adopting different atrous rates. The proposed RSIG module uses a resolution-matching submodule to enlarge an input feature map and a ResNeXt block to exploit richer semantic information. In verification experiments, the proposed 2CFF-SSD model achieved 66.41% average precision (AP) in the Oxford hand test set conducted at 55 frames per second on a GTX 1080 Ti graphics processing unit, which is superior to the original SSD method and many other state-of-the-art methods in terms of accuracy, while maintaining the high speed.