AI-powered licence plate recognition (LPR) — also known as automatic number plate recognition (ANPR) — reads vehicle registration plates from live camera feeds in real time, matching them against access control lists or watchlists within milliseconds. It is one of the most concrete, high-value applications of computer vision in physical security because it produces an immediately verifiable output: a plate number tied to a vehicle identity.
LPR serves four primary functions in physical security: automated access control at gates and barriers, parking management and enforcement, tracking vehicles of interest against watchlists, and evidence capture for post-incident investigation. In each case, the system replaces manual checks with automated, real-time verification.
The LPR pipeline consists of five stages, each building on the previous one. Understanding these stages helps integrators design systems that maximise read rates.
Before the system can read a plate, it must first detect a vehicle in the frame. The AI model identifies vehicles as distinct objects, drawing a bounding box around each one. This stage filters out non-vehicle motion — pedestrians, shadows, debris — ensuring the OCR engine only processes relevant frames.
Within the vehicle's bounding box, a second model (or a specialised layer within the same model) locates the licence plate region. This is a precise crop operation: the system identifies the rectangular area containing the plate characters, isolating it from the vehicle body, bumper, and background.
Plate localisation must handle significant variability: plates can appear at different sizes depending on distance, at various angles depending on camera mounting, and with different aspect ratios across countries and plate types.
The localised plate image is passed to an OCR (Optical Character Recognition) engine that reads individual characters. Modern AI-based OCR outperforms traditional template-matching OCR because it can handle variable fonts, character spacing, dirt and wear on plates, partial occlusion from tow bars or plate frames, and non-standard plate formats.
The OCR engine outputs a string of characters along with a confidence score for each character and for the plate as a whole. Characters with low confidence can be flagged for manual review rather than being silently misread.
The recognised plate string is immediately compared against one or more databases: an allow list (authorised vehicles), a deny list (banned or revoked vehicles), or a watch list (vehicles of interest). The lookup is near-instantaneous, typically completing in under 50 ms.
Based on the match result, the system triggers an action: open a barrier, deny access, or flag the vehicle for operator review.
The final stage is the physical or digital outcome. For access control, this means sending a signal to a barrier controller. For surveillance, it means generating an alert with the plate image, the recognised string, the match result, and the timestamp. In parking management, it may log the entry time for billing purposes.
Legacy LPR systems relied on rigid template matching: the system expected plates to appear in a specific size, font, and position within the frame. Any deviation — an angled approach, a dirty plate, a non-standard format — caused a failed read.
AI-based LPR handles variability by design. Deep-learning models trained on millions of plate images across diverse conditions can read plates through motion blur (vehicles travelling at speed), partial occlusion (tow bars, plate frames, dirt), non-standard and foreign plates, low-light and IR-illuminated conditions, and wide capture angles (up to 30° from perpendicular).
The key metric for LPR systems is the read rate — the percentage of plates successfully read — versus the false read rate — the percentage of plates read incorrectly. AI-based systems routinely achieve read rates above 95% with false read rates below 1%, compared to 70–85% and 3–5% for legacy OCR systems.
LPR is one of the most placement-sensitive applications of computer vision. Integrators must pay close attention to several variables to achieve reliable read rates.
Capture angle: The camera should be positioned at no more than 30° from perpendicular to the vehicle's direction of travel. Greater angles distort the plate characters and reduce recognition accuracy.
Resolution per plate: The plate must occupy a minimum number of pixels in the frame — typically at least 130 pixels across the width of the plate for reliable character recognition. This determines how far from the camera the capture zone can be.
IR illumination: For night operation, dedicated IR illuminators are recommended. Camera-integrated IR is often insufficient because the plate's reflective coating can cause overexposure. External IR at a controlled angle avoids this.
Mounting height: A height of 1.0 to 1.5 metres — roughly plate height — is ideal for straight-on captures at gates and barriers. Higher mounts require a steeper camera angle, which compresses characters.
Single vs dual camera: High-throughput lanes (motorways, busy car parks) often use a dedicated LPR camera paired with an overview camera. The LPR camera is optimised for plate capture; the overview camera provides context (vehicle colour, make, driver presence).
SafetyScope's Omni platform includes a dedicated LPR module that integrates with access control systems and watchlist databases. The system supports multi-country plate formats and handles the full pipeline — vehicle detection, plate localisation, character recognition, and database lookup — within a single deployment.
Alerts for watchlist matches are routed through the same alert pipeline as other detection types, appearing in the operator's VMS/PSIM dashboard alongside intrusion alerts, loitering events, and other security events. This unified approach means operators do not need a separate system for vehicle monitoring.
Published: 2025-11-17 · Updated: 2026-04-02