Deploying computer vision models in production presents unique challenges, particularly around processing speed and data privacy. Running high-throughput video analytics in the cloud incurs extreme bandwidth costs and latency, making edge processing the preferred path for modern physical security systems.
In this recorded session, Monika Sharma shares practical insights from scaling edge-based computer vision. She details how to compress neural networks using quantization, convert models to run on TensorRT, and process multiple RTSP streams without frame drops.
Privacy is also a major focus. We walk through the architecture of a real-time privacy pipeline that redacts faces and vehicle information at the ingestion point, ensuring the video streams are GDPR compliant before they are stored or processed for security metrics.