Back to Portfolio
Computer Vision

Sports Ad Tech

< 30ms Processing Latency

Real-time Engagement

In the world of live sports, every millisecond counts. This project involved building a high-speed computer vision pipeline capable of detecting brand logos and on-field assets in real-time to trigger contextual ad insertions and analytics.

Core Tech Stack

OpenCV / YOLOv8 CUDA / TensorRT C++ Core Modules AWS Media Services Redis Pub/Sub

Challenge

Analyzing live 4K streams with multiple fast-moving objects usually requires significant compute overhead, making real-time interactive ad-triggering prohibitively expensive and slow.

Solution

We optimized a lightweight YOLO-based model and deployed it on NVIDIA T4 GPUs using TensorRT. Our custom C++ bridge reduced the overhead between the video decoder and the inference engine, achieving sub-30ms latency.

Result

The system successfully analyzed 100% of live broadcast frames without dropping quality. This enabled our client to offer "dynamic contextual ads" for the first time, increasing ad revenue by 25% during major tournaments.