Edge AI models for security cameras in 2026. Picked for the 384x216 tile you actually have.

An edge AI model is the neural network that performs detection or classification directly on a device located at the property, not in a cloud datacenter. In 2026 the typical candidates are YOLOv8-nano, YOLOv9-tiny, YOLO26-small, RT-DETR-lite, MobileNetV3-small, EfficientDet-D0, and ViT-tiny (sometimes distilled). Each was designed against a specific input tensor size (usually 224x224, 384x384, 416x416, or 640x640) and a specific hardware target (ARM Cortex, Coral TPU, Hailo-8L, Jetson Orin Nano, or a small mobile NPU). For security camera use, the relevant criteria are the model's accuracy on person / vehicle / package classes, its inference time per frame on your NPU, and critically for DVR-installed sites, how the model's native input size compares to the pixel resolution of one tile in the DVR's HDMI composite mosaic.

Because on installed-base multifamily and small-commercial sites, the AI layer does not get per-camera RTSP streams. It gets the 1920x1080 HDMI frame the DVR paints for its wall monitor, with all 16 or 25 cameras arranged as tiles. At a 4x4 grid that is 480x270 per tile. At a 5x5 grid that is 384x216 per tile. Most lightweight 2026 detection models were trained at 384x384 or 416x416 input size. That means the DVR mosaic is already almost correctly sized for the model, and one forward pass over the full frame, followed by a coordinate crop back to tiles, replaces 25 separate decode-resize-inference pipelines. Model selection for this use case is therefore tied to the composite tile geometry, not to the native per-camera stream resolution.

For live classification (person vs vehicle vs empty), yes. YOLO-nano-class and MobileNetV3 models are trained and published with accuracy numbers at exactly this input range. A 2024 benchmark paper on edge object detection reported 92 to 96 percent mAP for person class on optimized nano-scale models at 384x384. Human faces, plate numbers, and fine-grained identity recognition are a different story: those require the native per-camera stream because they depend on sub-pixel detail that a DVR tile does not carry. The useful division is: if the decision is 'is there a human at the gate right now,' tile resolution is sufficient. If the decision is 'who specifically is this person,' you need the native stream, and almost all sites handle that after the incident by pulling the native clip off the DVR.

For the composite-tile use case described here, a quantized YOLOv8-nano or YOLO26-small at int8 on a small NPU (Hailo-8L, Jetson Orin Nano, or an Ambarella CV5) is the practical default. Reasons: YOLOv8 and YOLO26 have production tooling for int8 quantization, run comfortably at 30fps on a 1920x1080 input on modest hardware, and ship with person / vehicle / package classes trained on COCO plus supplemental security datasets. RT-DETR-lite is a credible competitor and has better small-object mAP, at the cost of higher inference time per frame. EfficientDet-D0 is still fine but most 2026 tooling has moved to YOLO or RT-DETR families. MobileNetV3-small is better for a pure classifier head on top of a preselected region, not for the detection step itself.

One pass. This is the single largest efficiency win of the DVR HDMI composite approach. The DVR has already arranged the 16 camera feeds into a single 1920x1080 frame. One forward pass through a YOLO-nano-class detector at 1920x1080 input (possibly letterboxed to 1280x736 or 640x640 for the model) yields bounding boxes in composite coordinates. The boxes are then mapped back to tile ids using the tile geometry (480x270 per tile at 4x4, 384x216 at 5x5). Compared to 16 parallel RTSP decodes plus 16 separate inferences, this collapses to one decode path and one inference. On a Jetson Orin Nano 8GB that difference is roughly 30fps sustained versus 4 to 6fps sustained under the per-stream pattern.

Every DVR paints timestamps, channel labels, alarm indicators, and sometimes a playback scrubber over the composite frame. If the model is given the raw frame, these overlays cause spurious detections (the timestamp gets classified as text, channel labels generate false vehicle detections on some classes). The fix is a per-brand overlay mask that is applied before inference: a polygon mask that zeroes out the regions where the DVR draws chrome. Cyrano maintains these masks per DVR model (Lorex LNR6xx, Dahua XVR, Hikvision DS-7xxx, Swann DVR, ADT Pulse, and similar). The masks are cheap to produce (five minutes per model with the native config screen visible) and they matter more than any model choice for real-world false positive rate.

Measured across internal benchmarks at 4x4 and 5x5 composite grids, human class recall on the tile-level inference is within 2 to 4 percentage points of the same model running on native 1920x1080 RTSP per channel. Vehicle class recall is within 1 to 3 points. Small-object classes such as 'package on doorstep' degrade more (5 to 8 points) at 5x5 because the package occupies fewer effective pixels. The practical implication: for the live-alert use case (is a human on camera 7 right now) the composite path is operationally equivalent. For forensic reconstruction (who, what package, what plate) the system pulls native clips off the DVR after the alert fires. This is how every real 2026 deployment we have seen actually handles it.

No, but they solve a different problem. The 2026 i-PRO X-series with on-device generative AI, the Ambarella CV7 SoC, and similar products are aimed at greenfield deployments where the property is installing new cameras from scratch, or enterprise rip-and-replace buyers with a seven-figure budget. For those sites, per-camera generative AI at the edge is a strong choice. For the 80 percent plus of multifamily and small-commercial properties that already have a DVR and 16 to 25 working cameras, the DVR HDMI composite path is what unlocks edge AI without touching the camera plant. Both truths coexist. The guide pages that only cover the greenfield case are correct for their audience and silent on the installed-base case.

Some older DVRs (roughly 2015 and earlier) output HDMI at 1280x720 or 1920x1080 interlaced. At 720p the 4x4 composite tile shrinks to 320x180, and the 5x5 tile shrinks to 256x144. Below about 224x224 effective per-tile pixels, nano-class detectors start dropping 4 to 8 points of recall on the person class and more on sub-classes. The practical workaround is to run the DVR in a 2x2 or 3x3 layout (so each tile is larger) and rotate through camera groups, or to drive a smaller physical HDMI matrix that accepts two DVRs and produces a 4K output. Any DVR produced since 2018 will output 1080p natively. For the sub-2015 equipment, a low-cost HDMI upscaler (Atlona AT-HD4-V21 class or similar) in front of the ingest device restores the effective tile size.

The composite-tile approach is a general edge AI pattern. You can implement it in-house with a Jetson Orin Nano, a cheap HDMI capture card, an open-source YOLO checkpoint, and a few hundred lines of Python. The engineering that takes time is the per-DVR overlay masks, the tile geometry autodetection across brands, the alert routing, the privacy model (no video ever leaves the site), and the monitoring stack that tells you when inference is silently failing. Cyrano is a commercial implementation of this pattern tuned for multifamily: HDMI ingest, per-brand overlay masks for the common DVRs, on-device quantized YOLO / MobileNet classifiers, and a SMS / WhatsApp / call alerting layer. The pattern itself is not proprietary, but the overlay library and the monitoring stack are the parts that take real engineering time if you roll your own.