AI-Driven Tracker Cuts Identity Switches by 42% in Synthetic Scenes

Corvus ISR tracker benchmark matrix (seed 1337) — The published matrix — every row reproducible. Source: corvusisr.com/benchmark

In the realm of wide-area motion imagery (WAMI), maintaining accurate object identities over time is a major challenge. Corvus ISR has just published a public benchmark showcasing two different models designed to improve multi-object tracking performance. These tests use synthetic scenes with perfect ground truth, providing a clear view of the trackers’ capabilities without real-world noise.

The baseline model, v1, employs a simple greedy nearest-neighbour approach with constant-velocity predictions and fixed 2-second coasting, making it a reliable yet basic reference point. The latest version, v2, introduces an auction-based tracker that uses a three-tier association process, velocity-consistency gating, and confidence-decayed coasting to enhance tracking accuracy. This more advanced approach aims to reduce identity errors, which are critical in surveillance and security applications.

Results from the benchmark reveal a significant reduction in identity switches: for example, in a scenario with 150 movers at 2 frames per second, switches dropped from 2,042 to 1,183 per minute — a 42.1% decrease. Similar improvements are seen with increased densities and degraded conditions, such as dense scenes with 400 objects, where switches decreased from 14,032 to 8,040, or roughly 42.7%. These numbers demonstrate the robustness of v2 even under challenging circumstances.

It’s important to note that the ID switch metric used here is intentionally strict, counting every change, fragmentations, and re-acquisitions as switches. This makes the results a genuine measure of tracker performance, not just a marketing claim. Corvus ISR publishes these numbers openly, emphasizing their commitment to transparency: the public benchmark provides the raw data, and you can even reproduce it live in your browser without signup or NDA.

From an engineering perspective, v2 runs efficiently, averaging around 1.2 milliseconds per sensor tick at high density, well within real-time constraints. This performance is achieved without hardware acceleration — just a sophisticated algorithm running in-browser, making it accessible to anyone interested in multi-object tracking technology. The entire setup is synthetic, generated pixel-by-pixel, ensuring perfect ground truth and eliminating variability from real-world factors.

By publishing these detailed failure numbers, Corvus ISR demonstrates a commitment to objective measurement over marketing hype. Despite improvements, even the best models still make thousands of identity errors per minute under stress, highlighting the ongoing challenge in this field. Future trackers will be tested against the same benchmark seed, ensuring a level playing field.

If you’re curious about how these advanced tracking algorithms perform in practice, you can directly test them yourself. Head over to the demo site and run the benchmark — see the real-time results firsthand and explore the cutting-edge of surveillance technology.