Model reference — Pictures · Video · Live feed
Crowd counting estimates how many people are in a scene, from a handful in a room to tens of thousands in a stadium or square, without identifying any of them.
Two methods apply at different densities. Up to a few dozen people, a detector finds each person individually and counts the boxes, which is accurate and also gives their positions. Beyond that, individuals overlap too heavily to be separated, and density estimation is used instead: the model produces a map of how many people occupy each part of the image and adds it up. Because nobody is identified, counting carries much lighter obligations than recognition, which makes it the appropriate choice for occupancy, safety and flow measurement.
Models in this group take a single image as input: a photograph, a scan or a screenshot. Each specification table gives three hardware tiers — Minimum, the smallest setup on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is the number of sample inputs processed per hour on that hardware. Use these rates for initial sizing. Before production, benchmark your own data to validate accuracy, latency, throughput and cost. The sample input here is one 2-megapixel image with the number of people counted.
Vendor: Tencent Youtu Lab
What it does: locates every individual as a point, which gives an accurate count and their positions even in dense crowds. The best general choice where the layout of the crowd matters.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 6,000 | ≈ 24,000 | ≈ 65,000 |
Vendor: University of Illinois
What it does: produces a density map for very dense scenes and totals it. Well established, light to run and dependable at stadium and street scale.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 12,000 | ≈ 48,000 | ≈ 140,000 |
Vendor: Stony Brook University
What it does: a density method with better calibration than earlier models, meaning its totals are closer to correct without per-camera adjustment.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 9,000 | ≈ 36,000 | ≈ 100,000 |
Vendor: Nanjing University
What it does: a transformer-based counter that is more accurate at awkward camera angles and severe perspective, where density methods over- or under-count at the far end of a scene.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 8 | 12 | 24 |
| RAM | 32 GB | 48 GB | 96 GB |
| Server | 1× RTX 3090 24 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 3,000 | ≈ 12,000 | ≈ 32,000 |
Vendor: Ultralytics
What it does: detects each person individually, which is more accurate than density estimation up to a few dozen people and also returns where they are. The right choice for rooms, queues and small gatherings.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 40,000 | ≈ 160,000 | ≈ 450,000 |
Models in this group take a recorded video file as input and are applied frame by frame. The three hardware tiers mean the same as above. The sample input here is one minute of 1080p video at 25 frames per second.
Vendor: Tencent Youtu Lab
What it does: counts people on each sampled frame of recorded video, producing an occupancy figure over time rather than a single number.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 240 | ≈ 900 | ≈ 2,400 |
Vendor: University of Illinois
What it does: the lighter option for long recordings, allowing every frame of hours of footage to be counted at modest cost.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 500 | ≈ 1,800 | ≈ 4,800 |
Vendor: Huazhong University of Science and Technology
What it does: follows individuals between frames, which turns a series of counts into flow measurements — how many distinct people passed, not how many were present.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 1,200 | ≈ 4,000 | ≈ 10,000 |
Models in this group take a live camera or stream as input and must keep pace with it in real time. The three hardware tiers mean the same as above, but the rate here is how many streams one server of that tier can keep up with in real time rather than a per-hour count. The sample input here is one 1080p stream at 25 frames per second.
Vendor: University of Illinois
What it does: holds real time on a live stream with little hardware, which is what allows continuous occupancy monitoring across many cameras.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (camera streams handled at 25 frames per second) | ≈ 8 streams | ≈ 30 streams | ≈ 90 streams |
Vendor: Tencent Youtu Lab
What it does: the accurate live option where crowd positions as well as totals are needed, for example to detect a dangerous build-up in one part of a space.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (camera streams handled at 25 frames per second) | ≈ 4 streams | ≈ 16 streams | ≈ 48 streams |
Vendor: Ultralytics
What it does: counts people live in rooms and doorways at low density, running comfortably on modest hardware alongside other detection on the same stream.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (camera streams handled at 25 frames per second) | ≈ 10 streams | ≈ 36 streams | ≈ 100 streams |
The method follows from your crowd density, camera angle and whether you need positions or only a total. Our consultants review your cameras and scenes, then recommend a model and the calibration needed to make the counts accurate on your views.
At the start of a project we may run a short proof of concept on a sample of your own data, measuring the accuracy and the throughput the model actually achieves on your material. That replaces the estimates on this page with real figures, so the cost and the schedule for the full engagement are known before it is committed.
Share a representative sample, expected volume, latency target and deployment location for Crowd counter. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster.