TortTeam

Our inspiration came from the massive gap between powerful, research-grade computer vision models and the reality of deploying them in production. Real-time video analytics often require expensive hardware, which makes them inaccessible for many edge devices or budget-friendly cloud setups. We wanted to bridge this gap by building a highly optimized, end-to-end pipeline that leverages the OpenCV AI ecosystem and AWS infrastructure to deliver real-time object detection and tracking without breaking the bank or sacrificing frame rates. TortTeam is a real-time video analytics pipeline that detects and tracks objects in live video streams. It ingests video feeds, runs optimized inference using OpenCV, and streams annotated results back to a user-friendly dashboard. By combining lightweight model architectures with OpenCV's highly optimized DNN module and deploying on AWS, the system achieves high frames-per-second (FPS) with minimal latency. Users can monitor live streams, view historical analytics, and receive alerts based on specific object detections. We built the core computer vision pipeline using Python and OpenCV, utilizing the cv2.dnn module for inference and optical flow algorithms for object tracking. The backend is powered by FastAPI to handle video streaming via WebSockets, while the frontend dashboard was built with React. We containerized the application using Docker and deployed it entirely on AWS infrastructure, leveraging EC2 instances for compute and S3 buckets for storing video logs, model weights, and processed artifacts. One of our biggest challenges was optimizing the video decoding and inference pipeline to avoid frame drops and memory leaks. We had to implement asynchronous processing and smart frame-skipping to ensure the UI remained responsive under heavy loads. Managing network bandwidth for live streaming was another hurdle; we solved this by dynamically adjusting the stream resolution based on the client's connection speed. Additionally, configuring the AWS IAM roles and VPC settings to securely allow our EC2 instances to communicate with S3 buckets took significant time and debugging. We are incredibly proud of achieving a stable 30+ FPS processing speed on a modest AWS instance, proving that high-quality CV analytics don't always require massive GPUs. We also successfully implemented robust object tracking that maintains identities even when objects are temporarily occluded. Finally, building a fully functional, end-to-end system that handles live video ingestion, processing, and a responsive UI within the tight hackathon timeframe is a huge accomplishment for our team. We learned a tremendous amount about the intricacies of video processing—specifically how to handle memory buffers and avoid bottlenecks when reading frames from live webcams and RTSP streams. We deepened our understanding of OpenCV's DNN module and how to properly pre-process frames for optimal inference. On the cloud side, we gained hands-on experience with AWS deployment strategies, containerization, and real-time data streaming at scale. Moving forward, we plan to integrate more advanced computer vision models, such as pose estimation and multi-camera tracking. We also want to optimize our pipeline to run directly on edge devices (like Raspberry Pi or NVIDIA Jetson) using AWS IoT Core. Finally, we aim to add a serverless architecture using AWS Lambda for on-demand video processing to further reduce cloud costs and improve scalability.

Built With

• amazon
• amazon-web-services
• api
• computer
• deep
• detection
• docker
• ec2
• fastapi
• learning
• numpy
• object
• opencv
• python
• pytorch
• react
• rest
• s3
• tracking
• vision
• websockets
• yolo