🤖 AI Detection

Object detection models, classes, and configuration for OKITO.

🎯 Models

Detection Model

Default model for object detection and pose estimation.

detection_model.pt

• ✓ Fast inference
• ✓ High accuracy
• ✓ GPU accelerated

Pose Detection Model

Pose estimation model for action detection.

pose_model.pt

• ✓ 17 keypoints COCO
• ✓ Sitting/standing detection
• ✓ Static pose phone

Sliced Detection

Sliced detection for small objects on high-res cameras.

sahi

• ✓ Slice: 128px
• ✓ Overlap: 0.6
• ✓ Small Area Ratio: 0.08

📋 Detection Classes

Available Classes

Class	Description	Default Threshold
person	Human detection	0.45
car	Car, truck, bus	0.45
cell phone	Phone detection	0.50
smoking	Smoking detection	0.70
on_phone	Phone at ear	0.65
phone_in_hand	Phone in hand	0.60

🎚️ Sensitivity Modes

Configuration

Mode	Description	Use Case
High	Minimum false positives, high precision	Critical security, production
Medium	Balanced mode (recommended)	General monitoring
Low	Maximum detections, may have false positives	Testing, low-risk areas

Recommendation: Use Medium for production. Switch to High only if you see too many false positives.

⚙️ Configuration

Detection Settings

Parameter	Type	Default	Description
model	string	detection_model.pt	Model file path
device	string	cuda:0	cuda:0 or cpu
imgsz	int	1536	Image size for inference
conf	float	0.45	Confidence threshold
iou	float	0.7	IoU threshold for NMS
classes	list	[person]	Classes to detect

Sliced Detection Settings

Parameter	Type	Default	Description
sahi_slice	int	128	Slice size in pixels
sahi_overlap	float	0.6	Slice overlap ratio
sahi_interval	int	2	Process every N frames
small_area_ratio	float	0.08	Small object area threshold