Local AI Person Detection Explained

When someone walks into your camera's field of view, AiSecam draws a bounding box around them in real time. This happens entirely in your browser — no frames are sent to any server. But how does a web page actually identify a person? The answer involves machine learning models, WebGPU acceleration, and some clever browser APIs.

The YOLO Model Family

AiSecam uses YOLO26n, a compact object detection model from the YOLO (You Only Look Once) family. YOLO models are designed for real-time detection: they analyze an entire image in a single pass through the neural network, producing bounding boxes and class labels for every object they find.

The "n" in YOLO26n stands for "nano" — this is a small, fast variant optimized for edge devices and browsers. Despite its small size, it accurately detects people, and it runs quickly enough to process live video at usable frame rates.

How the Detection Pipeline Works

Here is the pipeline step by step:

1. Camera access via WebRTC — the browser's getUserMedia API provides a live video stream from the device camera.
2. Frame extraction — each video frame is drawn onto an offscreen canvas, which converts it into pixel data the model can process.
3. Preprocessing — the image is resized to the model's input dimensions (typically 640x640) and normalized to the expected value range.
4. Neural network inference — the preprocessed image is fed through the YOLO26n model. This is where WebGPU acceleration makes a significant difference.
5. Post-processing — the raw model output contains hundreds of potential detections. Non-maximum suppression (NMS) filters these down to the most confident predictions, and a confidence threshold removes low-probability results.
6. Rendering — surviving detections are drawn as bounding boxes on a canvas overlay on top of the video feed.

WebGPU vs CPU Inference

The browser has two ways to run neural network inference: CPU (via WebAssembly) and GPU (via WebGPU). The difference in performance is dramatic.

WebGPU uses the device's graphics card to perform the massive parallel matrix multiplications that neural networks require. On a modern phone with a Snapdragon or Apple Silicon chip, WebGPU inference can process frames 3-5x faster than CPU-only inference. This means higher frame rates, more responsive bounding boxes, and lower battery drain.

CPU fallback — when WebGPU is not available (older browsers, some iOS versions), the model runs on the CPU via WebAssembly. This works but at reduced frame rates. On a mid-range Android phone from 2022, expect 5-8 FPS with WebGPU versus 2-4 FPS on CPU alone.

AiSecam automatically detects WebGPU support at startup and selects the best available backend. No configuration is needed.

Why Person Detection Stays Local

The model runs entirely in the browser's JavaScript execution context. The weights are loaded once when you start the camera and cached for subsequent sessions. Every inference happens on the device's processor — no pixel data is ever transmitted over the network.

This architecture has real privacy implications. Cloud-based detection services require uploading video frames to remote servers, where they are processed and the results sent back. Even if those servers are "secure," the act of transmission creates exposure. With on-device detection, the entire pipeline is self-contained within the browser sandbox.

Detection Classes

While YOLO models can detect 80 different object classes (cars, animals, furniture, etc.), AiSecam focuses on person detection to minimize false alerts. The confidence threshold is configurable — lower values make the detector more sensitive but increase the chance of false positives.

Performance on Different Devices

Detection performance varies significantly by device tier:

• Flagship phones (2023+) — 15-25 FPS with WebGPU, excellent bounding box tracking
• Mid-range phones (2021-2023) — 8-15 FPS with WebGPU, good real-time response
• Budget phones (2019-2021) — 3-8 FPS, usable but with noticeable latency
• Desktop browsers — 20-40 FPS depending on GPU, smooth detection

The practical threshold for useful person detection is around 3 FPS — below this, the bounding boxes lag too far behind the person's actual position to be helpful.