koreashin/Driver_monitoring

Driver Behavior Detection Model (Epoch 7)

운전자 이상행동 감지를 위한 Video Swin Transformer 기반 모델입니다.

Model Description

• Architecture: Video Swin Transformer Tiny (swin3d_t)
• Backbone Pretrained: Kinetics-400
• Parameters: 27.85M
• Input: [B, 3, 30, 224, 224] (batch, channels, frames, height, width)

Classes (5)

Label	Class	F1-Score
0	정상 (Normal)	0.97
1	졸음운전 (Drowsy Driving)	0.99
2	물건찾기 (Reaching/Searching)	0.96
3	휴대폰 사용 (Phone Usage)	0.96
4	운전자 폭행 (Driver Assault)	1.00

Performance (Epoch 7)

Metric	Value
Accuracy	98.05%
Macro F1	0.9757
Validation Samples	1,371,062

Training Configuration

Parameter	Value
Hardware	2x NVIDIA RTX A6000 (48GB)
Distributed	DDP (DistributedDataParallel)
Batch Size	32 (16 x 2 GPU)
Gradient Accumulation	4
Effective Batch	128
Optimizer	AdamW (lr=1e-3, wd=0.05)
Scheduler	OneCycleLR
Mixed Precision	FP16
Loss	CrossEntropy + Label Smoothing (0.1)
Regularization	Mixup (a=0.4), Dropout (0.3)

Files

File	Size	Description
pytorch_model.bin	121 MB	PyTorch weights (FP32)
model.onnx	164 MB	ONNX model for mobile deployment
config.json	1.2 KB	Model configuration
model.py	6.9 KB	Model architecture code
convert_coreml_macos.py	2.2 KB	CoreML conversion script (macOS)

Platform-specific Usage

PyTorch (Server/Desktop)

import torch
from model import DriverBehaviorModel

model = DriverBehaviorModel(num_classes=5, pretrained=False)
checkpoint = torch.load("pytorch_model.bin", map_location="cpu")
model.load_state_dict(checkpoint["model"])
model.eval()

iOS (CoreML)

1. Copy model.onnx to macOS
2. Run conversion script:

python convert_coreml_macos.py

3. Add generated DriverBehavior.mlpackage to Xcode project

Android (ONNX Runtime)

// build.gradle
implementation 'com.microsoft.onnxruntime:onnxruntime-android:1.16.0'

// Kotlin
val session = OrtEnvironment.getEnvironment()
    .createSession(assetManager.open("model.onnx").readBytes())

val output = session.run(mapOf("video_input" to inputTensor))

Preprocessing (All Platforms)

Input Shape: [1, 3, 30, 224, 224]  (batch, channels, frames, height, width)
Channel Order: RGB
Normalization: (pixel / 255.0 - mean) / std
  - mean = [0.485, 0.456, 0.406]
  - std = [0.229, 0.224, 0.225]
Resize: 224x224 (BILINEAR)
Frames: 30 frames uniformly sampled

Dataset

• Total Videos: 243,979
• Total Samples (windows): 1,371,062
• Window Size: 30 frames
• Stride: 15 frames
• Resolution: 224x224

Training Progress

Epoch	Accuracy	Macro F1
5	97.35%	0.9666
6	97.74%	0.9720
7	98.05%	0.9757

License

This model is for research purposes only.

Citation

@misc{driver-behavior-detection-2026,
  title={Driver Behavior Detection using Video Swin Transformer},
  author={C-Team},
  year={2026}
}