This study collected images and videos in which people were not correctly recognized by object recognition technology and investigated common characteristics of scenes where person detection became uncertain.
The investigation used the YOLO object recognition models, applying YOLOv8 to images and YOLOv5 to videos. Detection results and confidence scores were examined using dataset images, photographs taken by the researcher, and YouTube live camera footage. Scenes with unstable person detection included cases where faces or parts of the body were obscured, clothing colors were similar to the background, or lighting conditions involved darkness or backlighting.
In live camera footage, the same person carrying a parasol was identified as “person,” “bicycle,” and “motorcycle” at different moments. Other examples included a person crouching in a dark location being identified as a skateboard and a person whose upper body was obscured by an umbrella being identified as a sheep.
These results suggest that a combination of conditions related to a person’s appearance and surroundings may lead to misclassification or missed detections. Future work will investigate additional combinations of conditions to clarify when person detection becomes uncertain.
