Multi-object tracking (MOT) in wide-area motion imagery (WAMI) is a critical technology for surveillance, requiring accurate detection and consistent object identities over time. The challenge lies in minimizing identity switches, which occur when the system wrongly assigns a new identity to a tracked object. Corvus ISR’s latest benchmark demonstrates significant progress in reducing these errors, providing valuable insights for future development in the field.
The company published a public tracker benchmark comparing two models on an identical synthetic scene, with perfect ground truth to ensure measurement accuracy. The baseline, v1, uses a simple two-pass greedy association with fixed velocity prediction and coasting, establishing a performance floor that remains operational even in archived demo slices. In contrast, v2 introduces an auction-based tracker with track confirmation, velocity-consistency gating, and confidence-decayed coasting, aiming for smarter object association.
The results are impressive: in a scenario with 150 movers at 2 frames per second, the number of identity switches dropped from 2,042 to 1,183 per minute — a 42.1% reduction. For denser scenes with 400 movers, switches fell from 14,032 to 8,040, a 42.7% decrease. These metrics are strictly defined, counting every change in track identity, including re-acquisitions and fragmentations, making them a rigorous measure of tracker performance.
Beyond just accuracy, the benchmark highlights that even the advanced v2 tracker still commits thousands of identity errors under stress, emphasizing that perfect tracking remains a challenge. The synthetic scene setup with perfect ground truth ensures these numbers are true measurements, not marketing hype. Every future tracker will be publicly benchmarked against this same seed, fostering transparent progress in the field.
From an engineering standpoint, the v2 model achieves an average processing time of approximately 1.2 milliseconds per sensor tick at a density of 400 objects, with a worst-case of about 5ms — all within real-time browser performance. This accessible runtime means anyone can reproduce the benchmark results by simply visiting the live demo and pressing ‘Run benchmark’ — no signups or NDAs required. The development process was AI-driven, reviewed, and validated before release, ensuring trustworthy results.
All of this is built on a fully synthetic environment, with generated pixels and no real-world entities involved. This approach allows for precise measurement of tracker performance without extraneous variables, making the progress toward better multi-object tracking more transparent and measurable than ever before.
Interested in seeing how this benchmark performs in your own environment? Feel free to run the benchmark yourself and observe the improvements firsthand.
Automated Multi-Camera Surveillance: Algorithms and Practice (The International Series in Video Computing, 10)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
wide-area motion imagery system
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
GPS Tracker for Vehicles – No SIM Required, Hidden Magnetic Car Tracker, Real-Time GPS Tracking for Kids, Car, Pets, No Monthly Fee, Works with Apple Find My, 2026 Upgraded (Black)
【No Monthly Fee, No SIM, No Subscription Required】This mini GPS tracker for vehicles requires no SIM card, no…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Security Cameras Wireless Outdoor, 2K Indoor Cameras for Home Security Battery Powered, AI Motion Detection, Color Night Vision, 2-Way Talk, Spotlight Siren Alarm, Cloud & SD Storage-Jet Black Camera
2K HD Live Video, Picture & Color Night Vision: The security cameras wireless outdoor provide a degree wide…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
