Suspicious-Activity-Detector

The project pursues the development of Real-time Video Analysis-based Suspicious Activity Detector, which will utilize modern object detection algorithms in detecting anticipated shoplifting incidences and other activities that trespass the stores, with no assistance from the human operatives.

Approach

Technologies Used

1. YOLO (You Only Look Once): Utilized for real-time object detection, this deep learning model identifies and tracks objects in the video feed, drawing bounding boxes around detected entities such as individuals, bags, or merchandise.

2. TensorFlow: Facilitates video preprocessing and the development of machine learning pipelines, enabling seamless integration of object detection and behavioral analysis.

3. Pose Detection: Marks key body coordinates and identifies motion patterns, aiding in behavioral analysis to determine if movements align with suspicious activities.

4. Scikit-learn: Facilitates preprocessing and feature extraction from video data, supporting the classification and anomaly detection processes.

5. Convolutional Neural Networks (CNNs): Analyzes image and video frames for extracting spatial features, forming the backbone of object detection and pattern recognition tasks.

6. Pretrained EfficientNet Model: Utilized to extract robust features from video frames, this model accelerates development by leveraging transfer learning and fine-tuning for domain-specific tasks.

7. LSTM (Long Short-Term Memory Networks): Processes temporal information from video frames, analyzing sequences of actions to detect patterns indicative of suspicious behavior.

8. Reinforcement Learning with Human Feedback (RLHF): Enhances the system's ability to refine and improve its decision-making by incorporating feedback from store personnel on flagged incidents, ensuring the model learns from real-world
   scenarios and reduces false positives over time.

9. User Interface (UI): Combines a Python-based backend server and a frontend dashboard built using HTML and CSS. The UI processes video feeds, applies ML model inference, and displays annotated videos with bounding boxes and real-time behavioral labels. It also delivers alerts to store personnel and provides actionable insights via an intuitive dashboard.