shirish9354/Criminal_Activity_Detector_DaudshahVideoDataset

🎥 Criminal Activity Detection Model Card

Overview

This project presents a deep learning–based video classification system that categorizes short clips as Normal or Suspicious. The model is trained on the Daudshah Video Dataset and is designed for surveillance analysis, supporting both offline video processing and real-time streams.

The pipeline combines spatial feature extraction (CNN) with temporal modeling (RNN + attention) to capture motion patterns indicative of anomalous behavior.

---

Intended Use

✅ Primary Use Cases

• Surveillance monitoring (CCTV analysis)
• Smart city safety systems
• Automated alerting for abnormal activities
• Pre-screening large volumes of video footage

⚠️ Limitations / Out-of-Scope

• Not suitable for legal or forensic decision-making without human review
• Not intended for identifying individuals or facial recognition
• Performance may degrade under:
  • Low-light conditions
  • Unusual camera angles
  • Domains outside the training distribution

---

Model Architecture

Input Video (2-second clip, 16 frames)
             ↓
ResNet50 (ImageNet pretrained, frozen)
             ↓
Frame Embeddings (16 × 2048)
             ↓
LayerNorm → Dense(256) → Dropout(0.4)
             ↓
Bidirectional GRU (64 units)
             ↓
Temporal Attention Layer
             ↓
Dense(64) → Dropout(0.6)
             ↓
Sigmoid Output → P(Suspicious / Normal)

---

Key Design Choices

• Frozen CNN backbone → reduces overfitting on limited data
• Bidirectional GRU → captures forward and backward temporal dependencies
• Attention mechanism → emphasizes important frames within each clip

---

Training Details

Parameter	Value
Framework	TensorFlow / Keras
Backbone	ResNet50 (frozen)
Sequence length	16 frames
Clip duration	~2 seconds
Batch size	32
Epochs	~19 (early stopping applied)
Optimizer	Adam (clipnorm = 1.0)
Loss	Focal Loss (α = 0.60, γ = 2.0)
LR Schedule	ReduceLROnPlateau
Early Stopping	Based on validation AUC

Training Strategy

• Precomputed CNN features cached for efficiency
• Stratified video-level splits to prevent data leakage
• Decision threshold tuned using macro-F1 score

---

Dataset

Source

• Daudshah/Video-dataset
• ~7 GB worth of 20-30 seconds of surveillance video clips

Classes

• Suspicious (13 categories): Abuse, Arrest, Arson, Assault, Burglary, Explosion, Fighting, Road Accident, Robbery, Shooting, Shoplifting, Stealing, Vandalism
• Normal: Non-criminal surveillance footage

Data Splits

Split	Clips	Normal	Suspicious
Train	1,328	51.2%	48.8%
Validation	332	51.2%	48.8%
Test	290	51.7%	48.3%

---

Performance

Metric	Score
Accuracy	95.24%
AUC-ROC	0.9895
Macro F1	0.903

Class-wise Performance

Class	Precision	Recall
Normal	89.6%	92.0%
Suspicious	91.2%	88.6%

Error Analysis

• False Positives: 12 / 150
• False Negatives: 16 / 140

---

Inference

Supported Inputs

• Video files (MP4, AVI, etc.)
• Live streams (Webcam, RTSP, RTMP)

Inference Method

• Sliding window approach over video streams
• Default 50% overlap to reduce missed detections

Performance

• ~135 ms per 2-second clip (NVIDIA T4 GPU)
• Real-time capable with batching and warm-up

---

Known Issues & Mitigations

Issue	Mitigation
Class imbalance bias	Focal Loss tuning
Overfitting	Frozen backbone
Slow preprocessing	Batched GPU feature extraction
Threshold bias	Macro-F1 optimization
Serialization issues	Keras serializable registration
Training interruptions	Checkpointing with resume support
Feature extraction crashes	Incremental checkpointing
Dataset path issues	Auto-discovery logic
Data leakage risk	Video-level splitting

---

Ethical Considerations

• False positives may trigger unnecessary alerts
• False negatives may miss critical incidents
• Should always be used with human oversight
• Avoid deployment in sensitive or high-risk scenarios without proper bias evaluation

---

Future Work

• Fine-tune CNN backbone with regularization
• Extend to multi-class classification (crime type detection)
• Explore transformer-based temporal models
• Optimize for edge deployment (TensorRT / ONNX)
• Improve domain adaptation for real-world CCTV environments

---