Household robots are accelerating the evolution of smart-home living, and effective cleaning depends heavily on reliable perception and navigation. This study improves image-recognition capabilities within a robotic cleaning system by focusing on its image-processing module. The proposed approach refines image preprocessing and feature extraction to suppress noise and enhance robustness, and integrates regional stereo-matching constraints to achieve more accurate region-level correspondence. Dynamic obstacles are tracked in real time using a SURF–KLT scheme, and a Greedy strategy is then applied to pinpoint moving targets, lowering collision risk and increasing cleaning coverage. Experimental results show an image-matching accuracy of 99.7%, with an mAP@0.5 of 0.932 and stable training precision–recall behavior. In practical cleaning tests, the robot successfully identified 23 pieces of household waste and computed a weighted total score of 39 to support optimal path planning.
