Smart Home Target State Detection Method Based on Image Semantic Understanding

Abstract

This application discloses a method for detecting the state of smart home targets based on image semantic understanding. The method includes: acquiring an input image containing a smart home target and preprocessing it; obtaining a device body mask and target category through a device body instance segmentation model; calling the corresponding closed boundary template according to the target category to generate a closed boundary centerline, a closed boundary band, and inner and outer regions. Further, a foreign object mask is obtained through a foreign object semantic segmentation model, and connected component analysis is used to identify the foreign object's connected components. Cross-boundary evidence crossing the closed boundary centerline is determined based on the inner and outer regions and the foreign object's connected components. Simultaneously, the continuity and local obstruction features of the closed boundary are calculated within the closed boundary band, and the state label of the smart home target is determined accordingly. This method can identify pseudo-closed states caused by foreign object clamping, improving the accuracy and reliability of state detection.

Images