Skip to main content

Real-time Detection System for Suspicious Stabbing Movements

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

                  ┌────────────────┐
                  │  Video Stream  │
                  └───────┬────────┘
                          ▼
            ┌───────────────────────────┐
            │   Multi-Stage AI Model    │
            │  (YOLOv11 + Pose + LSTM)  │
            └─────────────┬─────────────┘
                          ▼
             /─────────────────────────\
            <   Suspicious Movement?   >
             \─────────────────────────/
                │ Yes             │ No
                ▼                 ▼
         ┌─────────────┐   ┌─────────────┐
         │ Save Face & │   │   Render    │
         │ Alert Logs  │   │ Normal Feed │
         └─────────────┘   └─────────────┘

Overview​

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

Unlike simple frame-by-frame classifiers, our system uses a Long Short-Term Memory (LSTM) neural network to evaluate keypoint movement vectors over a temporal buffer, allowing it to understand the context of movements over time.

Key Features​

  • Real-Time Analysis: Process local video feeds or camera streams with minimal latency.
  • Dual Stream Detection:
    • Object Detection (YOLOv11): Specifically fine-tuned to identify weapons (e.g. knives) with high precision.
    • Pose Estimation (YOLOv11-Pose): Tracks skeletal keypoints (shoulders, elbows, wrists) of multiple persons in the frame.
  • Temporal Sequence Modeling: Uses a Bidirectional LSTM network tracking a buffer of up to 150 frames of normalized keypoints to analyze arm swing speeds and stabbing motion dynamics.
  • Smart Frame Skipping: Implements dynamic frame skipping (processing 1 frame every $N$) to optimize performance on CPU/GPU hardware while maintaining the underlying keypoint sequence history.
  • Automatic Alert System:
    • Automatically crops and saves the face of the suspicious/violent person in a dedicated folder (suspect/).
    • Records detailed timestamped logs (logs/) with person IDs, action labels, and detection confidence values.
  • Interactive UI: Fully featured graphical interface built with Tkinter, featuring live feed rendering, dynamic frame skipping adjustment, and quick folder shortcuts.

Tech Stack​

The application is built on top of a modern machine learning and computer vision stack:

  • Programming Language: Python 3.8+
  • Computer Vision:
    • OpenCV for image manipulation and video capture.
    • Ultralytics YOLO11 for object detection (weapons) and human pose tracking.
  • Deep Learning:
  • Data Engineering:
    • Roboflow for dataset preparation and annotation.
    • NumPy & SciPy for vector normalization and math operations.
  • Graphical User Interface:
    • Tkinter for a lightweight, cross-platform dashboard.

Demo Videos​

You can view the system in action on different test cases (both violent actions and normal daily tasks) on YouTube:

🎥 Suspicious Stabbing Movements Detection Playlist

The playlist shows several scenarios:

  1. Aggressive Stabbing Detection: Red warning boxes and alerts triggered immediately.
  2. Normal Interactions: Green boxes highlighting daily hand movements and normal walking, proving low false-positive rates.
  3. Weapon Detection: Blue bounding boxes highlighting visible knives.