A Deep Dive on Creepy Cameras (hackaday.com)

Benn Jordan’s teardown and hands-on investigation into Automated License Plate Readers (ALPRs) exposes both how pervasive ML surveillance has become and how brittle those systems are. He outlines the typical ALPR pipeline—image segmentation/object detection to find plates, then OCR to read characters—and traces its U.S. deployment history. Jordan disassembles a Motorola in-vehicle ALPR and then shows how inexpensive edge hardware (a Raspberry Pi 5 + Halo AI board) running YOLO object detection can produce a far more accurate, low-cost ($~250) computer-vision stack than many commercial law‑enforcement units. The video’s most consequential findings are practical adversarial countermeasures: a transparent sticker pattern that keeps plates human-readable while confusing ALPR detectors (causing misreads or missed detections) and other low-tech tricks like IR or even yarn-based interference. Technically this underscores that plate detection often relies on brittle heuristics (e.g., filtering for rectangles) and that physical-world perturbations can defeat segmentation+OCR pipelines. For the AI/ML community this is a reminder that edge deployment, dataset bias, adversarial robustness, multi-frame aggregation, and sensor fusion need stronger attention—because both surveillance and effective countermeasures are increasingly cheap, open-source, and easy to prototype (code available on Jordan’s GitHub).