• 

• 

  Shoplifting Tren over the years

• 

  Preprocessed videos

• 

  Preprocessed videos

ShopLift-VLM: Privacy-Preserving Shoplifting Detection

Inspiration

Shoplifting remains a persistent threat to retail, with U.S. retail theft losses reaching $112.1 billion in 2022, escalating to an estimated $121.1 billion in 2023, and projected to exceed $143 billion by 2025. Despite these staggering figures, only about 2% of shoplifters are apprehended. Traditional video surveillance systems generate vast amounts of data that security personnel cannot analyze in real time, opening the door for AI-powered systems that enable rapid detection, alerting, and valuable insights.

What It Does

• Privacy-Preserving Detection: Blurs faces in real time to protect identities.
• Context-Aware Analysis: Leverages vision-language models to interpret complex behavioral cues.
• Edge Compatibility: Optimized for resource-constrained devices (e.g., Raspberry Pi 4, Jetson Nano).
• Real-Time Alerts: Flags potential shoplifting incidents immediately.

What's Next

• Dataset Expansion: Increase data diversity from various retail environments.
• Real-Time Monitoring: Integrate with live security feeds for instant alerts.
• Hardware Optimization: Adapt for additional accelerators like Coral TPU and Jetson Orin.
• Behavioral Refinement: Enhance detection of subtle, complex shoplifting tactics.
• User Interface: Develop an interactive dashboard for security monitoring.
• Field Testing: Pilot the system in real-world retail settings.

Built With

• 4-bit-quantization
• datasets
• dcsass-dataset
• google-colab-pro
• google-gemma-4b/1b-vision
• groq
• jetson-nano
• jupyter-notebooks
• matplotlib
• numpy
• nvidia-gpus
• onnx-runtime
• opencv
• pandas
• pil/pillow
• pixtral-12b
• python
• pytorch
• qlora-fine-tuning
• raspberry-pi
• shoplift-vlm-dataset
• tensorrt
• tqdm
• transformers
• trl
• unsloth
• yolov11