~$1.55
Per 1K Images (AWS)
Rekognition PPE detection
13 TOPS
Edge AI (Hailo-8L)
Pi 5 PCIe accelerator
PPE + Count
Detection Types
Hard hat, hi-vis, worker count
Bolt-on
Deployment Model
Add to existing DSLR-Pi fleet
AI / CV Options
| Platform | Capability | Cost | Notes |
|---|---|---|---|
| AWS Rekognition | Built-in PPE detection | ~$1.55 / 1K images | Zero setup — API call per image |
| Google Vision AI | Label detection, AutoML for custom models | ~$2.33 / 1K (labels), ~$5 / 1K (AutoML) | AutoML requires training data |
| YOLOv8 Self-Hosted | Real-time object detection | Free (GPU server required) | Full control, custom training |
| Construction-Specific | Newmetrix (Procore), viAct, Eyrus (AU) | Enterprise pricing | Purpose-built for construction safety |
Edge AI Hardware
| Platform | Performance | Cost | Notes |
|---|---|---|---|
| Pi 5 + Hailo-8L | 13 TOPS | $250–350 | Official Pi support, PCIe M.2 HAT |
| Pi 4 + Coral USB | 4 TOPS | $190–220 | Simpler setup, well-documented with TFLite |
| Jetson Orin Nano | 40 TOPS | $350–450 | Best raw AI performance, NVIDIA ecosystem only |
| Pi 5 + Hailo-8 (full) | 26 TOPS | $270–400 | Double the Hailo-8L, same PCIe interface |
What AI Can Detect
PPE Compliance
- Hard hat, hi-vis vest, safety boots detection
- AWS Rekognition has this built in
- Edge: YOLOv8 custom-trained on PPE dataset
- Accuracy degrades in dust, rain, partial occlusion
Worker Counting
- Person detection + tracking across frames
- YOLOv8 + ByteTrack on edge hardware
- Zone-based counting for shift reports
- Useful for productivity and labour tracking
Progress Tracking
- Compare current frame to baseline image
- Detect slab poured, framing up, windows installed
- Structural change detection via image differencing
- Requires well-defined milestone definitions
Safety Zones
- Restricted area intrusion detection
- Geofenced zones with real-time alerts
- Crane swing radius, excavation perimeters
- Requires fixed camera position for calibration
Recommended Approach
Don't build a new camera system. Bolt AI onto the existing DSLR-Pi fleet. Start with AWS Rekognition PPE detection as an optional add-on. Graduate to edge AI (Pi 5 + Hailo-8L) when volume justifies it.
AI accuracy in real construction conditions (dust, rain, partial occlusion) is unreliable. False positives create alert fatigue. Liability risk if AI says “compliant” and someone gets hurt.
Competitors
| Company | Focus | Notes |
|---|---|---|
| Newmetrix (Procore) | AI safety observations | Integrated with Procore platform |
| viAct | Construction safety AI | Purpose-built, works with standard IP cameras |
| Eyrus (Australian) | Workforce management | Zone monitoring using existing CCTV |
vs DSLR-Pi: This is a software overlay, not a replacement. Adds $250–450 for edge hardware or ~$1.55/1K images for cloud.