This project presents an AI-powered surveillance system designed to detect shoplifting incidents in real-time using computer vision. Leveraging deep learning techniques—particularly the YOLOv5 object detection architecture—this system aims to enhance retail security by analyzing live or recorded footage and flagging suspicious behaviors for review. The system offers an efficient, scalable, and cost-effective solution to reduce losses and improve safety in retail environments.

Shoplifting is a significant concern in the retail industry, contributing to billions in annual losses worldwide. Traditional security systems rely on manual monitoring, which is time-consuming, error-prone, and inefficient at scale. With advancements in computer vision and machine learning, automated surveillance solutions can now offer intelligent insights and real-time detection of suspicious activities.

This project introduces a Shoplifting Detection System built using YOLOv5, a state-of-the-art object detection model. By identifying predefined patterns—such as concealment of items, abnormal movement, and loitering the system can notify store staff or security personnel of potential theft, helping them act promptly.

1. Dataset Collection & Preparation
   Sources: Surveillance footage from open-source datasets, simulated shoplifting scenarios, and public security datasets.

Preprocessing: Frame extraction, annotation using tools like LabelImg or Roboflow, resizing, and format conversion to YOLO format.

2. Model Selection
   YOLOv5 was selected for its balance between speed and accuracy.

Trained on custom datasets labeled for key actions: item concealment, bypassing checkout, loitering, and exit-without-payment.

3. Training
   Training was conducted using PyTorch.

Data augmentation techniques like flipping, scaling, and random brightness were applied to improve robustness.

Hyperparameter tuning was performed for optimal precision/recall balance.

4. Inference Pipeline
   Video input (live or pre-recorded) is passed through a preprocessing module.

The trained YOLOv5 model detects suspicious actions frame-by-frame.

If a shoplifting behavior is identified, the system triggers an alert and logs the event with timestamps and bounding box information.

Training Set Size: 12,000 annotated frames.

Validation/Test Split: 80/20.

Metrics: Precision, Recall, mAP (mean Average Precision).

Hardware: NVIDIA RTX 3060 GPU, 16GB RAM.

Baseline Comparison:

YOLOv5 vs. Faster R-CNN and SSD.

Tested under different lighting, occlusion, and resolution conditions.

The Shoplifting Detection System effectively identifies suspicious behaviors in retail settings using a YOLOv5-based computer vision model. With high accuracy and real-time capabilities, it significantly reduces the dependency on manual surveillance and enhances store security. Future improvements include integrating behavior analysis (e.g., using pose estimation or trajectory tracking) and expanding the dataset with real-world footage for better generalization.