Safety detection checks whether the protective equipment a site requires is actually being worn — hard hat, high-visibility vest, safety glasses, gloves, ear defenders, harness — and whether people are where they should not be: inside an exclusion zone, under a suspended load, on the wrong side of a barrier.
PPE stands for personal protective equipment. The technical task is detection plus association: it is not enough to find a helmet and a person in the same frame, because the helmet may be on a bench or on someone else. A working system attaches each item to the person wearing it, which needs the person’s body position as well as the object. What determines whether such a system is used, though, is the alert rate. A camera watching a busy site produces tens of thousands of judgements an hour, and at any realistic accuracy that means a stream of alerts nobody can act on unless the rules are tight: a condition must persist for a set number of seconds, in a defined zone, during defined hours, before anything is raised.
Models in this group take a live camera as input and must keep pace with it in real time so that an unsafe condition can be acted on while it exists. 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 how many camera streams or feeds one server of that tier can keep up with in real time, not a per-hour count: live work must fit inside the interval between frames, and a server that cannot keep pace drops frames rather than falling behind. 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 1080p stream at 25 frames per second.
YOLO11
Vendor: Ultralytics
What it does: detects people and PPE items on every live frame, fitted to the specific equipment your site mandates. The usual backbone of a live safety system, in several sizes so accuracy can be traded against how many cameras one server carries.
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
RT-DETR
Vendor: Baidu (PaddlePaddle)
What it does: produces steadier detections without a duplicate-removal step, which reduces the flickering that causes false alerts when a helmet is briefly obscured.
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)
≈ 5 streams
≈ 18 streams
≈ 50 streams
D-FINE
Vendor: University of Science and Technology of China
What it does: better on small and partly hidden objects — safety glasses, gloves, a harness clip at distance — which is where PPE detection on wide yard views usually fails.
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
≈ 15 streams
≈ 42 streams
RTMPose
Vendor: Shanghai AI Laboratory (OpenMMLab)
What it does: locates each person’s head, hands and torso, which is what attaches a detected helmet or glove to the right person rather than to whoever stands nearby.
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)
≈ 6 streams
≈ 22 streams
≈ 60 streams
ByteTrack
Vendor: Huazhong University of Science and Technology
What it does: follows each person through the scene so a violation is one alert for one person, rather than an alert on every frame they remain in view.
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)
≈ 12 streams
≈ 40 streams
≈ 110 streams
YOLO-World
Vendor: Tencent AI Lab
What it does: detects an item from a written description without being fitted first, which lets a newly mandated piece of equipment be enforced the day the rule changes rather than after a labelling exercise.
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)
≈ 3 streams
≈ 12 streams
≈ 34 streams
Input type — Video
Models in this group take recorded footage as input and are applied frame by frame, for after-the-fact auditing rather than intervention. The three hardware tiers mean the same as above, but the rate here is the number of sample inputs processed per hour rather than a stream count. 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 minute of 1080p video at 25 frames per second.
YOLO11
Vendor: Ultralytics
What it does: detects people and PPE across recorded footage, which is how a site is audited after the fact — compliance rates by area, by shift, by contractor — rather than alerted in the moment.
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)
≈ 1,600
≈ 6,000
≈ 16,000
RT-DETR
Vendor: Baidu (PaddlePaddle)
What it does: the steadier detector for audit work, where consistency across long footage matters more than latency.
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)
≈ 700
≈ 2,600
≈ 7,000
RTMPose
Vendor: Shanghai AI Laboratory (OpenMMLab)
What it does: associates equipment with the person wearing it across the recording, and additionally identifies unsafe postures such as lifting with a bent back or reaching into a machine.
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)
≈ 900
≈ 3,200
≈ 8,500
BoT-SORT
Vendor: Tel Aviv University
What it does: holds a person’s track through occlusion so a compliance figure counts distinct people rather than reappearances.
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)
≈ 800
≈ 2,800
≈ 7,000
Qwen2.5-VL 7B / 72B
Vendor: Alibaba Cloud
What it does: judges a flagged frame in context — distinguishing a helmet carried from a helmet worn, or a barrier crossed legitimately from a breach — which is what cuts the false-alert rate on the frames another model flagged. Applied only to those frames, never to all of them.
Requirement
Minimum
Medium
High
GPU type
RTX 4090 (7B, reduced precision)
A100 80 GB (7B, full precision)
2× H100 80 GB (72B model)
VRAM
16 GB
80 GB
160 GB combined
vCPUs
8
16
32
RAM
32 GB
64 GB
200 GB
Server
1× RTX 4090 24 GB
1× A100 SXM 80 GB
2× H100 SXM 80 GB
Rate (video minutes processed per hour)
≈ 40
≈ 160
≈ 90 (72B model, higher accuracy)
SAM 2 (Segment Anything 2)
Vendor: Meta
What it does: outlines a flagged person or zone precisely, which is used where a distance must be measured — how close someone stood to an edge or a load — rather than merely whether they were present.
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 (video minutes processed per hour)
≈ 120
≈ 450
≈ 1,200
Choosing between them
Requirements differ between a live intervention system and an after-the-fact safety audit, and between a fixed gate and an open yard. Our consultants review your camera coverage, your PPE rules and the alert volume your supervisors can act on, then recommend the models, the zones and the confirmation rules — and, where the purpose is evidence rather than intervention, how long footage should be kept and who may see it.
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 Safety / PPE detection. 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.