Securing CCTV Cameras Against Blind Spots

Overview

CCTV systems have "blind spots" where detection confidence is lower due to factors like angle and distance. This talk explores a novel issue with object detection: location-based confidence weakness. We show how a pedestrian’s position impacts detection confidence and analyze this across four lighting conditions and five object detectors (YOLOv3, Faster R-CNN, SSD, DiffusionDet, RTMDet). Footage from Broadway, Shibuya Crossing, and Castro Street reveals consistent blind spots. We introduce TipToe, an evasion attack that exploits these blind spots to create a minimum-confidence path, reducing detection confidence by up to 0.16 in Shibuya Crossing with YOLOv3 and similar results elsewhere.