Detection / segmentation
YOLO-family pipelines for human, package, defect, and general object detection.
For doorbells, IPC, industrial inspection, and service robots, Shunter delivers mass-production edge AI vision algorithms: data collection, model training, quantization, NPU / SoC on-board deployment, and production regression—shortening the cycle where “the demo works but hardware is hard to ship.”
Algorithms and ISP / sensor chains are tuned by the same team to avoid “accurate model, unstable picture” split deliveries.
System flow
From data to on-board: full train-deploy chain
A mass-production algorithm delivery path split into six coordinated stages from assets through launch.
-
01
Data collection
Scene samples
Annotation standards -
02
Model training
YOLO / MobileNet
Detection and classification pipelines -
03
Quantization
INT8 quantization
Operator adaptation -
04
On-board deployment
NPU / SoC inference
Latency and power tuning -
05
Imaging bring-up
With ISP / sensor
Joint validation -
06
Production regression
Version control
Continuous iteration
Solution stack
Four layers working together
Lightweight backbones
MobileNet and others for ARM / low-power cameras—classification and feature tasks.
Quantization and deployment
INT8 quantization, operator adaptation, on-board inference, performance and power tuning.
Scene customization
Iterate on customer data and conditions—sustainable, upgradeable model assets.
Implementation path
Six steps from requirements to mass production
Typical algorithm projects advance by phase so demo capability reaches firmware and volume.
-
1
Metrics and class definition
Lock detection classes, accuracy, false-alarm rate, latency, and power budget.
-
2
Data and annotation standards
Collect real-scene samples; unified annotation and QA process.
-
3
Training and evaluation
Select backbone, train, validate on holdout sets, analyze failure cases.
-
4
Quantization and board adaptation
INT8 quantization, operator replacement, inference on target NPU / SoC.
-
5
Full product bring-up
ISP, event policy, streaming logic—validate real doorway / line conditions.
-
6
Production regression and iteration
Model versioning, sampling regression, continuous fine-tuning.
Reference case
Delivery example: human / package recognition model SHT-A201
The train-deploy chain as a mass-production pre-trained model pack for locks, access control, and IPC.
Product capabilities
- Pre-trained pack for human, face, package, and related events on locks, access control, and IPC
- Continue fine-tuning on customer data for regional and scene differences
- Deliverable with YOLO general detection and MobileNet lightweight packs
How this solution maps to the delivery
- Pre-trained baseline: On-device models ready for demo and pilot
- Quantized deployment: INT8 and performance tuning for target SoC / NPU
- Full product coordination: Bring-up with lock / peephole / IPC imaging—control false alarms and power
Customer value
- Shorter path from algorithm validation to firmware integration
- Same capability extends to industrial defect, traffic object, and other model packs
- Combines with AI Vision Solution and PCBA for a complete product loop
More scenarios
Extensions of the same train-deploy capability
Security IPC
Zone intrusion and abnormal behavior assist—reuse doorbell detection assets.
Industrial AOI
Defect detection model packs—customer sample retraining and on-board regression.
Robot perception
Target follow and obstacle assist—combinable with multi-camera / TOF solutions.
Request a consultation AI Vision Solution Visual PCBA Customization