This paper explores the "edge cases" of computer vision: tracking objects that are both incredibly small and move with chaotic, high-speed trajectories. While modern tracking algorithms excel at following pedestrians or cars, this study reveals how they stumble when faced with a racquetball moving at 150 mph.
TL;DR
Researchers evaluated five state-of-the-art Kalman Filter-based trackers on a new, high-speed racquetball dataset. They discovered that while DeepOCSORT provides the best accuracy, all current methods suffer from significant "tracking drift" because their underlying linear motion models cannot keep up with the erratic, non-linear bounces of tiny, fast-moving objects.
What Problem Does This Paper Address?
In the world of Multi-Object Tracking (MOT), we’ve become quite good at tracking "predictable" things. If you are tracking a person walking across a street, a linear motion model (like a standard Kalman Filter) works beautifully because people don't instantaneously change direction at 100 mph.
However, in sports robotics or drone surveillance, we face fast-moving tiny objects. A racquetball is a nightmare for computer vision: it’s small (limited visual features), it moves incredibly fast (causing motion blur), and it bounces off walls in non-linear patterns that defy simple physics assumptions. The authors tackle the gap between standard tracking benchmarks and these high-intensity, real-world scenarios.
Methods
The authors employed a tracking-by-detection framework. This involves two steps: first, an object detector identifies the ball in every frame, and second, a tracker links those detections over time to form a trajectory.
- The Detector: After comparing YOLOv5, YOLOv7, YOLOv8, and Faster R-CNN, the team selected YOLOv8 for its superior balance of precision and inference speed on tiny objects.
- The Trackers: They benchmarked five Kalman Filter (KF) based algorithms:
- SORT: The baseline "Simple Online and Realtime Tracking."
- ByteTrack: Known for its "low-confidence" association logic to handle occlusions.
- OCSORT & DeepOCSORT: These focus on "observation-centric" recovery, helping the tracker "re-find" the object after it’s been lost or moved erratically.
- BoTSORT & StrongSORT: Advanced variants that incorporate camera motion compensation and deep appearance features.
The core of the analysis focuses on how these trackers handle the predict-update cycle when the "prediction" (where the KF thinks the ball will be) is constantly invalidated by a sudden bounce.
Datasets
The study introduces a custom racquetball dataset:
- Scale: 10,000 annotated frames.
- Resolution: 720p to 1280p, captured via a ZED 2i stereo camera.
- Complexity: Four distinct scenarios involving high-speed bounces and ricochets off sidewalls.
- Availability: The project code is hosted on GitHub.
Key Insights and Ideas
- The "Linear" Trap: The fundamental weakness is the Kalman Filter’s assumption of linear motion. When a racquetball hits a wall, its velocity vector changes instantly. The KF tries to "smooth" this, leading to a lag where the tracker thinks the ball is still moving forward while it has actually bounced back.
- The Speed-Accuracy Paradox: ByteTrack was the fastest (26.6ms inference), but it had the highest error. Because it prioritizes speed and simple association, it "drifts" significantly when the ball's path becomes complex.
- Appearance vs. Motion: In this domain, motion cues are often more reliable than appearance. Because the ball is so tiny (often just a few pixels), "Deep" features (like those in StrongSORT) don't provide as much benefit as they do for larger objects like humans.
- Cascading Failures: If the YOLO detector misses the ball for even two frames due to motion blur, the KF’s internal state breaks, leading to fragmented trajectories.
Key Results
- Accuracy Champion: DeepOCSORT achieved the lowest tracking error with an Average Displacement Error (ADE) of 31.15 pixels.
- The Performance Gap: Even the best tracker (DeepOCSORT) had error rates 3-4x higher than what is typically seen on standard MOT benchmarks.
- Spatial Drift: Trackers exhibited spatial errors ranging from 3cm to 11cm. For a ball that is only about 5.7cm in diameter, an 11cm error means the tracker is effectively "missing" the ball's actual location.
- Inference Speed: Most trackers maintained near real-time performance (~24-32ms per frame), but accuracy varied wildly across the four scenarios.
Limitations and Future Work
The authors are refreshingly honest: current Kalman Filter-based approaches are inadequate for precise high-speed tiny object tracking. The linear motion assumption is a "hard ceiling" for performance.
Future Directions: 1. Physics-Infused Models: Instead of a generic KF, trackers could incorporate a "physics engine" that understands gravity, air resistance, and the geometry of the court to predict bounces. 2. Hyperparameter Tuning: Standard trackers are tuned for pedestrians; the authors suggest specialized tuning for high-velocity objects. 3. Hybrid Systems: Combining deep learning-based trajectory prediction with traditional filters to better handle non-linear "snaps" in movement.
Summary updated