Challenges and Solutions for Efficient, Large-Scale Video Analysis

Processing a 50-minute video shouldn’t take 45 minutes yet that’s exactly where the detection part of our pipeline began, a pipeline consisting of detection, tracking and identification. When your goal is to analyse years of footage across dozens of cameras, that kind of performance isn’t just slow; it makes the entire project impossible. It would take us over a year to complete the run of our pipeline.

The task involved processing four years of historical footage from twelve cameras, with the expectation that the pipeline will later scale to over 100 live cameras. This presentation will show you the steps we took that transformed that bottleneck and deployed it into a fast scalable pipeline without touching the original machine learning models. Instead, most of the gain came from parallelisation and multithreading. We will also touch on other optimisations, some obvious and other that required some deeper knowledge and investigation. By implementing these techniques, we managed to get the -processing time of the detection part from 45 minutes for a 50-minute video down to two-three minutes.

We had the same goal through the entire project, optimise for speed and reduce cost of running the pipeline. This led us to investigate which virtual machine configuration that would deliver the best balance of cost and performance for our workload.

Along the way, we discovered how much performance can be unlocked simply by understanding the behaviour of the pipeline, from GPU saturation and I/O bottlenecks to the impact of video encoding choices. If you find yourself working on video analysis, real-time ML pipelines or largescale data processing, this session will show you our practical techniques and hard-won lessons, and how we squeezed more performance out of our system.