New AI-Powered Tracker Cuts Identity Switches by 42% in WAMI Tests

Understanding multi-object tracking (MOT) is crucial for wide-area surveillance, where systems monitor dozens to hundreds of moving objects across vast scenes. A key challenge in MOT is maintaining identity consistency — ensuring that each detected object is correctly identified over time, despite occlusions, noise, or clutter. The latest benchmarks from Corvus ISR demonstrate significant progress in this area, especially with their advanced tracker models.

The study compares two models: the baseline v1 ‘greedy nearest-neighbour’ approach, which uses a simple, fixed-velocity prediction method, and the newer v2 ‘confirmed-track auction’ system. The v2 model incorporates a sophisticated three-tier auction association process, velocity-consistency gating, and noise-scaled reservation pricing to improve tracking accuracy. This results in a remarkable 42% reduction in identity switches — from 2,042 to 1,183 per minute in a fixed test scene with 150 movers at 2 frames per second.

In dense scenarios with 400 movers, the improvements are equally impressive: switches drop from 14,032 to 8,040, a 42.7% decrease. Even under less ideal conditions—such as frame rates dropping to 0.5 fps, scenes with 20% occlusion, or degraded image quality—the v2 tracker maintains a significant edge, reducing errors by approximately 18% across these stress tests. These results are measured against perfect ground truth, with the system’s detection rate fixed by sensor properties, ensuring the comparison isolates tracker performance.

Notably, the metric honesty in this benchmark is strict: every change in track identity counts as a switch, including fragmentations and re-acquisitions, making these numbers a true reflection of system robustness. Corvus ISR publishes these results openly, emphasizing transparency — because every future tracker must face the same public benchmark, just like these.

From an engineering perspective, the v2 tracker operates at around 1.2 milliseconds per sensor tick at 400 density, well within real-time constraints. The entire process runs in the browser — no special hardware, no signup required. By using a fully synthetic environment with perfect ground truth, Corvus ISR ensures that these measurements are objective and reproducible.

The significance for tech enthusiasts and system integrators is clear: the auction-based tracker not only significantly reduces identity errors but also demonstrates that advanced algorithms can run real-time in a browser environment. If you’re interested in seeing these improvements firsthand, you can reproduce it live and run the benchmark yourself. This open approach invites transparency and shows how modern AI techniques are transforming wide-area surveillance systems today.

Powered by Thorsten Meyer AI