Real-time Detection System for Suspicious Stabbing Movements

An advanced real-time surveillance system designed to detect violence and potentially dangerous stabbing movements using Computer Vision and Deep Learning.

Overview

This project implements a multi-stage pipeline to identify violent intent in real-time video streams. By combining object detection, pose estimation, and temporal sequence analysis (LSTM), the system can distinguish between normal activities and suspicious stabbing motions.

Key Features

• Real-time Analysis: Processes video feeds with low latency.
• Object & Pose Detection: Simultaneously detects persons, weapons (knives), and body keypoints.
• Temporal Logic: Uses Long Short-Term Memory (LSTM) networks to understand the context of movement over time.
• Optimized Performance: Implements frame skipping logic (processing 1 frame every N) while maintaining the required input buffer for the LSTM model.

---

Demo

Test videos of the system detecting suspicious stabbing movements can be seen here:

https://www.youtube.com/playlist?list=PLo1f0U_Wr1t2QE8hyGbKqB8WCAeb5DOUH

The playlist includes multiple real-world test scenarios showing both violent and non-violent actions processed by the detection pipeline.

---

System Architecture

The workflow consists of four main stages:

1. Object Detection (YOLO11)

   • Identifies instances of person and knife.
   • Fine-tuned model: Specifically trained to recognize knives with high precision using a custom dataset.
   • Associates detected weapons with the closest person/hand.

2. Pose Estimation (YoloPose)

   • Extracts skeletal keypoints (wrists, elbows, shoulders).
   • Tracks motion vectors to analyze arm dynamics.

3. Human Activity Recognition (HAR)

   • Model: LSTM (Long Short-Term Memory).
   • Logic: Classifies motion patterns based on a sequence of 150 frames to detect stabbing actions.

4. Alert System

   • Triggers a real-time alert when the confidence threshold for violent action is exceeded.
   • Saving the face cropepd from the frame of the person who confidence threshold for violent action is exceeded

---

---

Tech Stack

• Core: Python, OpenCV
• Computer Vision: Ultralytics YOLO (Object Detection & Pose)
• Deep Learning: TensorFlow / Keras (LSTM)
• Data Management: Roboflow (Dataset preprocessing)
• Datasets:
  • Custom Knife Dataset
  • Automatic Violence Detection Dataset (GitHub)
  • Indoor action DAtaset (GithHub)
  • Real Life Violence Situations Dataset (Kaggle)
  • RWF2000 (Kaggle)

---

Getting Started

Prerequisites

• Python 3.8+
• CUDA-enabled GPU (recommended for real-time inference)

Installation

1. Clone the repository

   git clone [https://github.com/lraton/real-time-violent-action-detection.git](https://github.com/lraton/real-time-violent-action-detection.git)
   cd real-time-violent-action-detection/src/

2. Install dependencies

   pip install -r requirements.txt

Running the Application

Option A: Windows (CPU / Standard Shell)

Use this for testing without specific GPU configuration.

python main.py

Option B: WSL / Linux (GPU Accelerated)

Recommended for best performance using TensorFlow with GPU support.

# Activate your environment (if applicable)
source ~/my_venv/bin/activate

# Run the main script
python main.py