English | Tiếng Việt

Note: AI-assisted coding tools were used for minor implementation support

Violence Detection System with YOLO26 + Temporal Classifier

Real-time violence detection on edge devices

• 50 FPS on Raspberry Pi 5
• 81% accuracy on RWF2000
• 0.88 ROC-AUC
• Cross-dataset evaluation (RLVS / Hockey / Movies)

Demo

Watch demo here: https://youtu.be/Z1gKG_AFHuk

Overview

This project presents a lightweight violence detection system designed for real-time deployment on CPU-based edge devices.

The system combines spatial detection and temporal analysis:

• YOLO localizes potential violence regions
• A temporal classifier analyzes frame sequences to detect violent behavior

The goal is to achieve reliable violence detection while maintaining high inference speed on low-power hardware such as Raspberry Pi.

Quick review

Metric	Score
Accuracy	81.25%
ROC–AUC	0.886
FPS(Raspberry Pi 5)	51 FPS

Features

• Spatial Detection: YOLO26 for real-time object detection and localization
• Multi-Object Tracking: Tracker with confidence-based hysteresis
• Temporal Classification: Temporal classifier on 8-frame feature sequences
• Robust Tracking: Automatic tracker failure recovery and confidence decay
• Real-time Performance: Optimized for CPU and GPU inference
• Adaptive Frame Sampling: Configurable detection intervals for efficiency

Quickstart / Installation

git clone https://github.com/NgMQuang/Lightweight-YOLO-based-real-time-violence-detection-on-CPU
cd Lightweight-YOLO-based-real-time-violence-detection-on-CPU

pip install -r requirements.txt

Download weights and place them inside:

ViolenceDetector/
├── violence_yolo.onnx
├── temporal_classifier.onnx
├── temporal_classifier.onnx.data
└── demo.mp4

Also place a demo.mp4 file in the folder for testing.

cd ViolenceDetector
python ViolenceDetection.py

Performance

Training and Evaluation

RWF2000(Val only)

Training metrics

• Classifier

Metric	Score
Accuracy	0.8125
Precision	0.7990
Recall	0.8350
F1 Score	0.8166
Specificity	0.7900
False Positive Rate (FPR)	0.2100
False Negative Rate (FNR)	0.1650
ROC–AUC	0.8861

• Suspicious area localization

Class	Images	Instances	Box(P - R - mAP50 - mAP50-95)
all	3000	2865	0.712 - 0.704 - 0.754 - 0.425

Testing

To evaluate generalization ability, the model was tested on other datasets WITHOUT TRAINING ON THEM.

Metric	RLVS	HKF	Movies
Accuracy	0.7655	0.8090	0.7463
Precision	0.7010	0.7388	0.8182
Recall	0.9260	0.9560	0.6300
F1 Score	0.7979	0.6620	0.7119
Specificity	0.6050	0.7900	0.8614
False Positive Rate (FPR)	0.3950	0.3380	0.1386
False Negative Rate (FNR)	0.0740	0.0440	0.3700
ROC–AUC	0.9037	0.9247	0.8574

Real-Life Violence Situation (Full)

---

HockeyFight(Full)

---

MovieFight(Full)

Raspberry Pi 5 Implementation

Average FPS: 51.74

Pipeline timing:

Task	Average(ms)	Min(ms)	Max(ms)
Detection	10.426	0.000	165.764
Tracking	2.118	0.000	14.342
Classifier	0.542	0.000	15.600
Visualization	0.683	0.284	43.064
Frame latency	19.327	4.006	171.141

Weights

https://drive.google.com/drive/folders/10E4KqX_fWGagm4lv79oJ9eFl63tIdKg7?usp=drive_link

or you can find in releases

Configuration

FPS_VIDEO = 30                   # Video frame rate, auto read when assign video path
TOTAL_TIME_DETECT = 2.5          # Detection window (seconds), DO NOT CHANGE
FRAME_PER_DETECT = 8             # Frames per classifier input, DO NOT CHANGE
DETECT_INTERVAL = 10             # Automatically computed by the implementation.

TRACKER = "MEDIANFLOW"           # Tracker type, support MOSSE and KCF
MAX_TRACKS = 5                   # Max simultaneous tracks, DO NOT CHANGE
CONF_ON = 0.25                   # Show track threshold
CONF_OFF = 0.1                   # Hide track threshold
STICK_WEIGHT = 0.7               # Stickiness in scoring
alpha = 0.8                      # EMA smoothing factor
TRACKER_FAILURE_DECAY = 0.5      # Confidence decay on failure

How It Works

1. Detection Phase (Every N frames)

• YOLO detects suspicious areas
• Returns bounding boxes with confidence scores
• Extracts feature vectors

2. Tracking Phase (Between detections)

• Tracker updates box positions frame-to-frame
• Confidence scores decay if tracker fails
• Boxes with low confidence are removed

3. Classification Phase (Every N frames)

• Collects last 8 feature vectors
• Passes to temporal classifier
• Outputs violence probability (0.0 - 1.0)
• Alerts if probability > 0.8

4. Hysteresis & Display

• Tracks shown/hidden based on confidence thresholds
• Color-coded bounding boxes with track IDs
• Violence probability displayed on frame

Model Details

YOLO26 (violence_yolo.onnx)

• Input: 256×320 RGB images (normalized 0-1)
• Output:
  • Detections: (5, 6) - up to 5 boxes with [x1, y1, x2, y2, conf, class]
  • Features: (896, 15) - feature vector for temporal analysis
• Inference Time: ~50-200ms (CPU)

Temporal Classifier (temporal_classifier.onnx + .onnx.data)

• Input: (8, 896, 15) - 8 consecutive feature vectors
• Output: (2) - logit for binary classification [1] for fight and [0] for non-fight
• Inference Time: ~1-5ms (CPU)

Output

The script displays:

• Bounding boxes around detected people
• Track IDs and confidence scores
• Tracker status ("OK" or "HOLD")
• Violence probability when classified
• Alert when violence confidence > 0.8

Dataset

Models trained on custom dataset derived from RWF2000 (Real World Fighting Dataset):

• Contains real-world violence/non-violence scenarios
• Custom labeling
• 0.75 mAP on detection task
• 81.25% accuracy on violence classification

Dataset used for testing:

• Hockey Fight
• Movie Fight
• RLVS

Future Work

• Improve spatial feature extraction
• Optimize for embedded deployment

References

• YOLO26: https://github.com/ultralytics/ultralytics
• ONNX Runtime: https://onnxruntime.ai/
• RWF2000 Dataset: https://github.com/mchengny/RWF2000-Video-Database-for-Violence-Detection

License

MIT License

Acknowledgments

• YOLO26 by Ultralytics
• RWF2000 dataset creators
• ONNX Runtime community
• Hockey Fight
• Movie Fight
• RLVS