An Analysis of Kalman Filter based Object Tracking Methods for Fast-Moving Tiny Objects
Prithvi Raj Singh, Raju Gottumukkala, Anthony Maida · 2025
Read An Analysis of Kalman Filter based Object Tracking Methods for Fast-Moving Tiny Objects by Prithvi Raj Singh, Raju Gottumukkala, Anthony Maida (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-18451.txt.
Introduction
A racquetball moves fast, bounces in wild ways, and shows as a tiny mark on screen. Prithvi Raj Singh, Raju Gottumukkala, and Anthony Maida test how well classic trackers follow such an object in this 2025 paper. Tracking leans on Kalman filters, which predict motion step by step. Those filters do well on tame targets but strain here.
The team scored five top tracker variants on a custom set of 10,000 labeled racquetball frames. They weighed two factors: processing speed and update rate. One variant, DeepOCSORT, had the lowest tracking error by a wide margin. Another, ByteTrack, was a hair faster per frame. Yet all trackers drifted badly, with errors three to four times higher than on standard tests. The authors say fast, tiny targets need new methods built for them. Sports robots stand to gain from that work.
Catalog record
| CREATOR | Prithvi Raj Singh, Raju Gottumukkala, Anthony Maida |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-18451
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
An Analysis of Kalman Filter based Object Tracking Methods for Fast-Moving Tiny Objects Prithvi Raj Singh, Raju Gottumukkala, Anthony Maida Abstract Unpredictable movement patterns and small visual mark make precise tracking of fast-moving tiny objects like a racquetball one of the challenging problems in computer vision. This challenge is particularly relevant for sport robotics applications, where lightweight and accurate tracking systems can improve robot perception and planning capabilities. While Kalman filter-based tracking methods have shown success in general object tracking scenarios, their performance degrades substantially when dealing with rapidly moving objects that exhibit irregular bouncing behavior. In this study, we evaluate the performance of five state-of-the-art Kalman filter-based tracking methods—OCSORT, DeepOCSORT, ByteTrack, BoTSORT, and StrongSORT—using a custom dataset containing 10,000 annotated racquetball frames captured at 720p-1280p resolution. We focus our analysis on two critical performance factors: inference speed and update frequency per image, examining how these parameters affect tracking accuracy and reliability for fast-moving tiny objects. Our experimental evaluation across four distinct scenarios reveals that DeepOCSORT achieves the lowest tracking error with an average ADE of 31.15 pixels compared to ByteTrack’s 114.3 pixels, while ByteTrack demonstrates the fastest processing at 26.6ms average inference time versus DeepOCSORT’s 26.8ms. However, our results show that all Kalman filter-based trackers exhibit significant tracking drift with spatial errors ranging from 3-11cm (ADE values: 31-114 pixels), indicating fundamental limitations in handling the unpredictable motion patterns of fast-moving tiny objects like racquetballs. Our analysis demonstrates that current tracking approaches require substantial improvements, with error rates 3-4x higher than standard object tracking benchmarks, highlighting the need for specialized methodologies for fast-moving tiny object tracking applications. Project Code is avilable at https://github.com/Prithviraj97/KFTrackingEvaluation K eywords Tiny Object \cdot Object Tracking \cdot Kalman Filter \cdot Tracking Algorithms \cdot Racquetball 1 Introduction Recent advances in object tracking have led to significant improvements in both single-object tracking (SOT) and multi-object tracking (MOT) applications, with modern systems achieving less than 5% error rates on standard benchmarks involving larger objects. Modern tracking systems can effectively capture deep visual features of target objects, resulting in enhanced tracking performance across various scenarios. However, most of the progress around object tracking has been centered around tracking relatively large and stationary objects that follow predictable, linear motion patterns. Tracking rapidly moving small objects, such as racquetballs or tennis balls, is more challenging problem. Racquetball can exceed speed of 150 mph and can have multiple direction changes per second creating hard to predict non-linear trajectory pattern. The difficulty arises from several factors: the small size of objects leads to limited visual information for feature extraction, high-speed motion causes motion blur and rapid scale changes, background clutter can easily overwhelm small targets, and frequent occlusions make continuous tracking difficult. In the specific context of racquetball, these challenges are amplified because the ball exhibits multiple short-distance bounces and ricochets off sidewalls, creating highly erratic and non-linear motion patterns with sudden acceleration changes that are difficult to predict and track reliably over extended period. Object tracking approaches generally fall into two main categories: tracking-by-detection and tracking-without-detection. In tracking-by-detection methods, the primary focus is on robust object detection in each frame, followed by association and tracking of detected objects across all video frames. This approach typically combines deep learning-based object detection algorithms—such as Fast-RCNN [ 1 ] , Faster-RCNN [ 2 ] , YOLO models [ 3 , 4 , 5 ] , or SSD [ 6 ] — to locate target objects, with tracking algorithms deployed on top of the detection system to associate objects across frames while maintaining object identities and estimating future positions. In tracking-without-detection method, the tracking algorithm is solely responsible for tracking the target without performing detection at each frame. The algorithm itself estimates the future position, velocity, and other parameters of the target based on the given initial position and subsequently previously available information. Motion or object appearance based cues are used to track the target object. Optical flow based method [ 7 ] , mean-shift tracking [ 8 ] , and correlation filters [ 9 ] are some of the common algorithm for tracking-without-detection. However, these methods face particular challenges with tiny objects where there is limited appreance cues for feature extraction. For tracking-by-detection systems, various tracking algorithms can be employed, including Kalman filters, particle filters, and Intersection over Union (IoU) matching. All five object trackers examined in this paper (SORT [ 10 ] , DeepOCSORT [ 11 ] , OCSORT [ 12 ] , ByteTrack [ 13 ] , BoTSORT [ 14 ] , and StrongSORT [ 15 ] ) utilize Kalman filters as their core tracking mechanism, with modifications to initial state parameters and other configuration settings to improve performance. These methods assume linear motion models that become increasingly inaccurate as object motion becomes more erratic, leading to the performance degradation we observe in our experiments. Our research contributes to the field by providing a comprehensive analysis of Kalman filter-based tracking performance on fast-moving tiny objects. We address critical gaps in current tracking evaluation by introducing a novel racquetball dataset and specialized evaluation metrics. Specifically, we: • Present experimental results from implementing five state-of-the-art trackers on a novel racquetball dataset containing fast-moving tiny objects - the racquetball • Analyze the performance characteristics of these trackers using ADE and AMD metrics and identify key areas for improvement • Conduct thorough analysis of tracking limitations, identifying a critical trade-off where ByteTrack processes frames 15% faster (26.6ms vs 26.8ms average) but exhibits 260% higher tracking error compared to DeepOCSORT • Demonstrate that current kalman filter-based approaches show limitations for tiny object tracking, with all methods exhibiting substantial spatial errors that make them unsuitable for precise tracking applications • Provide detailed discussion of the architecture and working principles of Kalman filter and all five implemented tracking methods. 2 Related Works 2.1 Overview of Object Tracking Visual object tracking has evolved significantly over the past decades, transitioning from traditional computer vision methods to sophisticated deep learning approaches visual object tracking is an important area in computer vision, and many tracking algorithms have been proposed with promising results. Modern tracking approaches can be broadly categorized into generative trackers [ 16 ] , discriminative trackers [ 17 ] , [ 18 ] , [ 19 ] , [ 20 ] , and collaborative trackers [ 21 , 22 ] , with deep neural network-based methods gaining prominence due to their outstanding performance. Recently, object tracking algorithms based on deep neural networks have emerged and obtained great attention from researchers due to their outstanding tracking performance. The field has primarily focused on two main paradigms: tracking-by-detection and tracking-without-detection. In tracking-by-detection approaches, object detection is performed in each frame, followed by data association to link detections across frames. This methodology has proven effective for general object tracking scenarios but faces unique challenges when applied to tiny objects with rapid motion. 2.2 Small Object Detection and Tracking Small object detection and tracking represent particularly challenging areas within computer vision Object detection and tracking are vital in computer vision and visual surveillance, allowing for the detection, recognition, and subsequent tracking of objects within images or video sequences. The [ 23 ] survey paper highlights various approaches for small object detection and various benchmarks. The challenges arise