Why Computer Vision in Security Cameras Is More Limited Than the Ads Suggest

The current generation of consumer and prosumer security cameras all describe themselves in roughly the same terms: AI-powered detection, person recognition, package alerts, vehicle detection, facial recognition available with a subscription. The marketing implies a system that watches intelligently and tells you exactly what matters, filtering out the false alarms that have plagued motion-triggered cameras for decades.

The gap between that description and how these systems actually perform—in varied lighting, with complex backgrounds, with people who don’t cooperate with the camera’s viewing angle—is significant. The computer vision technology is real and genuinely useful. The specific claims made about accuracy, reliability, and capability range from reasonable to substantially overstated. Here’s what’s happening inside these systems and where the limitations come from.

What “AI Detection” Actually Means in 2026

Consumer security cameras use trained neural network models—typically convolutional neural networks (CNNs) or more recent transformer-based architectures—to classify objects detected in the video stream. When the camera’s motion sensor or pixel-change algorithm triggers on movement, the AI classification model runs on either the camera’s on-device processor or a cloud server, classifying the moving objects as “person,” “vehicle,” “animal,” “package,” or similar categories.

The key word is “classify.” The model looks at a frame (or a short sequence of frames) and produces probability estimates for object categories—essentially: “I think this is 78% likely a person, 12% likely an animal, 10% other.” The camera’s software then uses a threshold (often not visible to the user) to decide whether to send an alert. If the threshold is set low for sensitivity, you get more true detections but also more false positives. If it’s set high for precision, you miss some real events.

The models are trained on large datasets of labelled security camera footage—thousands or millions of examples of people, vehicles, and animals in various conditions. This training data mostly comes from cameras similar to those sold by the manufacturer, which means the model performs best on scenarios similar to what it was trained on and degrades on outliers.

The “AI” framing implies sophistication that the underlying system sometimes doesn’t have. Classifying a clearly visible adult human walking in front of a clean background in daylight is a solved problem that has been achievable since 2015. The hard cases—a child in winter clothing, a person at the edge of the frame in poor lighting, someone obscured by shadows or foliage—are where the classification accuracy drops significantly.

The Conditions Under Which It Works Well

Modern AI detection in security cameras is genuinely reliable under a specific set of conditions:

• Adequate lighting (daylight, or good artificial light coverage)
• Subject occupying a significant portion of the frame (not at the edge or very far away)
• Clean, relatively simple background without many plants moving in wind
• Normal-sized adult humans in recognisable clothing
• Subjects that aren’t moving extremely fast across the frame

In these conditions, the major camera systems (Nest, Ring, Arlo, Wyze, Reolink) achieve person detection accuracy in the 85–95% range in independent tests, with false positive rates (non-person movement triggering a person alert) in the 5–15% range depending on environment. This is genuinely useful—it dramatically reduces the alert volume compared to pure motion detection and catches most real events.

Night vision performance is more variable. Most cameras use infrared illumination for night vision, which produces a monochrome image that the AI models handle less accurately than colour daylight footage. The same models that hit 90%+ accuracy in daylight often drop to 70–80% accuracy in IR night mode, particularly for objects at the edges of the IR illumination range. Some cameras now use colour night vision (using wide-aperture lenses and image processing rather than IR), which produces better AI accuracy but doesn’t penetrate darkness as effectively.

The False Positive Problem Is Larger Than Reported

Security camera manufacturers report their AI detection accuracy in controlled benchmark conditions. These conditions are selected to show the system in a favourable light—adequate lighting, standard subjects, clear backgrounds. Real-world deployment produces significantly higher false positive rates because real environments don’t match benchmark conditions.

Common false positive sources:

• Foliage movement. Bushes and trees moving in wind create patterns that confuse zone-based motion detection and, more specifically, that produce shape contours the AI classifies as partially animal or human. A dense bush waving vigorously in strong wind generates alerts at a surprisingly high rate on many systems.
• Light changes. Cloud shadows moving across the scene, headlights of passing cars sweeping across the field of view, and sunrise/sunset angle changes all trigger motion detection and can produce false classifications.
• Small animals at camera range. A large dog at 5 metres is regularly classified as “person” by systems whose training data doesn’t strongly differentiate dog-sized quadrupeds from small humans. Squirrels and birds at close range trigger detections. Most systems have improved animal classification, but the boundary cases remain frequent.
• Bags, boxes, and stationary objects in motion. A plastic bag blowing across the driveway, a rolling trash can, or a slowly moving shadow from a flag are classified inconsistently—sometimes as vehicles, sometimes as people, sometimes as animals.

The practical effect is that users in environments with any of these features receive a substantial volume of false alerts. After enough false alerts, users either adjust sensitivity settings (reducing true positive capture), disable certain zone monitoring, or start ignoring alerts—which defeats the system’s purpose.

Facial Recognition: The Gap Between Marketing and Reality

Several camera platforms offer “familiar face” recognition—the ability to distinguish known individuals from strangers and alert differently based on the distinction. This feature is real in the limited sense that these systems maintain a database of enrolled faces and compare detected faces against it.

The accuracy of consumer-grade facial recognition in security camera contexts is substantially lower than what the marketing suggests. Consumer cameras have image resolution, compression artifacts, viewing angles, and lighting conditions that are significantly worse than the controlled frontal face images used to train and evaluate facial recognition systems. A face captured at 15 feet in late afternoon side lighting, partially in shadow, at a 45-degree angle from the camera is a much harder recognition problem than a frontal face in a well-lit indoor environment.

Published accuracy claims for consumer camera facial recognition typically don’t specify the conditions under which accuracy was measured. When tested in realistic deployment conditions—outdoor cameras at real house distances, with subjects who aren’t looking directly at the camera—recognition accuracy drops substantially. The “familiar face” feature is more useful as a secondary tool to investigate reviewed footage than as a real-time alert trigger.

Facial recognition in consumer security cameras also carries significant civil liberties implications that the marketing avoids entirely. Police in several US states have used Ring camera footage, and Amazon’s partnerships with law enforcement have been documented by the Electronic Frontier Foundation. The facial recognition database that “recognises your family” is also a database of biometric data held by a company whose cooperation with law enforcement requests is governed by terms of service rather than Fourth Amendment protections.

On-Device vs Cloud Processing: What It Actually Changes

A significant marketing distinction among current cameras is between on-device AI processing (running the model locally on the camera’s chip) and cloud-based processing (uploading footage to servers for analysis). The privacy and latency advantages of on-device processing are real.

On-device processing avoids sending footage to company servers before you see an alert. It works offline if your internet is down. It typically produces faster alerts because the classification happens on the camera rather than waiting for a round-trip to a server. Cameras with on-device processing from companies like Eufy, Reolink, and newer Arlo models with HomeBase offer these benefits.

The trade-off is that on-device AI runs on constrained hardware—the chips in security cameras are much less powerful than cloud GPU clusters. This limits model complexity and, typically, accuracy. Cloud-based processing can run larger, more accurate models on better hardware. The gap between on-device and cloud accuracy is narrowing as camera SoCs improve, but it hasn’t closed.

For privacy-conscious users, the on-device trade-off is clearly worth it. For users prioritising detection accuracy in difficult conditions, cloud processing from platforms with strong models (Google Nest, which uses Tensor AI infrastructure) may perform better. This is a real trade-off, not just a marketing distinction.

What Actually Helps Detection Performance

Camera placement is the variable that most dramatically affects AI detection performance—more than camera brand, model, or subscription tier. The ideal placement for reliable person detection:

• Camera mounted at 8–10 feet height, angled 15–45 degrees downward
• Subject walks toward the camera (not across it), occupying maximum frame percentage
• Minimal busy background behind the detection zone (no dense foliage, no road with frequent vehicle traffic)
• Detection zone configured to exclude areas with frequent false positive sources (specific trees, parts of road)

Most DIY security camera installations violate one or more of these principles because the ideal camera placement and the available mounting points on a house don’t always coincide. A camera that “has to” go in a corner because that’s where the electrical outlet is, or mounted higher than optimal because that’s the bracket location, performs meaningfully worse than a camera placed at the optimal detection geometry.

Supplemental lighting—a separate motion-activated light illuminating the monitored zone in addition to the camera’s built-in IR—dramatically improves night detection. A camera that performs poorly in darkness often performs well when the scene is illuminated by an external light, because the AI models perform significantly better on properly lit colour footage than on IR monochrome.

The Reasonable Expectations Calibration

A well-configured AI security camera system in a reasonable environment will catch most approaching people in the frame, produce alerts with a few minutes of lag on cloud systems or near-real-time on on-device systems, and give you useful footage to review when something happens. It will miss some events (poor lighting, unusual subjects) and generate some false alerts (environment-specific sources). It won’t reliably identify who a person is unless they’re directly facing the camera in good light. It won’t catch everything.

The marketing framing of “AI-powered” implies a system that intelligently understands what it sees. The reality is a system that classifies objects using statistical patterns from training data, with accuracy that varies dramatically with conditions. Understanding that distinction helps set expectations that match the technology, and helps make camera placement and configuration decisions that get the best out of systems that are genuinely useful—just not as omniscient as the packaging implies.