Physical Violence Detection Using Key Framing (Android + CNN)

Project Objective

Deliver an accurate, efficient video violence detection app that turns long videos into representative key frames and classifies each segment using a CNN. Designed for academic projects, research demos, and proof-of-concepts where explainability (frame evidence + timestamps) matters.

How It Works

1. Upload: Select a video from the gallery (MP4/AVI/MOV supported).
2. Key Framing: Motion- and scene-change–based frame sampling extracts representative frames per interval.
3. CNN Inference: Each key frame (or short clip) is resized/normalized and passed to a trained CNN (TFLite). Outputs class logits and confidence.
4. Aggregation: Frame-level predictions are temporally smoothed to create segment labels (fight, kick, punch, gunshot, none).
5. Report: Results with thumbnails, timestamps, and confidences are shown in-app and saved to Firebase for history.

Project Modules

Firebase Realtime Database Structure (Example)

{
  "users": {
    "uid123": {"name":"Alex","email":"a@b.com"}
  },
  "videos": {
    "vid_17000001": {
      "owner":"uid123",
      "path":"gs://bucket/videos/vid_17000001.mp4",
      "durationSec": 92,
      "createdAt": 1725345600
    }
  },
  "analyses": {
    "vid_17000001": {
      "frames": [
        {"t": 2.1, "img":"gs://bucket/frames/vid_17000001/f_0021.jpg", "label":"none",   "conf":0.81},
        {"t": 18.6,"img":"gs://bucket/frames/vid_17000001/f_0186.jpg", "label":"punch",  "conf":0.74},
        {"t": 36.2,"img":"gs://bucket/frames/vid_17000001/f_0362.jpg", "label":"fight",  "conf":0.83}
      ],
      "summary": {
        "topClass":"fight",
        "timeline":[{"start":16.0,"end":40.0,"class":"fight"}]
      }
    }
  }
}
              

Sample Firebase Security Rules (High-Level)

// Realtime Database Rules (illustrative)
{
  "rules": {
    ".read": false,
    ".write": false,

    "users": {
      "$uid": {
        ".read": "$uid === auth.uid",
        ".write": "$uid === auth.uid"
      }
    },

    "videos": {
      "$vid": {
        ".read": "data.child('owner').val() === auth.uid",
        ".write": "auth != null && ( !data.exists() || data.child('owner').val() === auth.uid )"
      }
    },

    "analyses": {
      "$vid": {
        ".read": "root.child('videos/'+$vid+'/owner').val() === auth.uid",
        ".write": "root.child('videos/'+$vid+'/owner').val() === auth.uid"
      }
    }
  }
}
              

Key Features & Benefits

• Explainable output with frame thumbnails and timestamps.
• Fast inference via key framing and on-device TFLite CNN.
• Configurable labels & thresholds for your dataset.
• Offline-first—queue logs and sync when online.
• Secure Firebase Auth + Rules for private projects.

Android Integration Sketch (Java/XML)

// PSEUDO-CODE (illustrative)

// 1) Pick Video
void pickVideoFromGallery(){
  // ACTION_OPEN_DOCUMENT with video/*, persistable URI permission
}

// 2) Key Framing
List<Frame> extractKeyFrames(Uri video){
  // Use MediaMetadataRetriever to sample frames every N ms
  // Apply motion/scene change heuristic to keep representative frames
  // Return list of (bitmap, timestamp)
}

// 3) CNN (TFLite) Inference
Prediction runModel(Bitmap frame){
  // Resize to model input (e.g., 224x224), normalize
  // Interpreter.run(input, output)
  // Map logits -> labels {fight, kick, punch, gunshot, none}
}

// 4) Temporal Smoothing
List<SegmentLabel> smooth(List<Prediction> preds){
  // Sliding window majority + confidence threshold
}

// 5) Firebase Save
void saveAnalysis(String vidId, List<FrameResult> frames, Summary summary){
  // Write to /analyses/{vidId} and upload thumbnails to Storage
}
              

What You Get