Womprat: Open-Source Small Aerial Target Tracking Without a GPU

A Tracking Solution That Doesn't Rely on Deep Learning

In the computer vision tracking space, mainstream solutions typically depend on GPU acceleration and pre-trained deep learning models. However, the team at Punk Labs recently open-sourced a tool that takes the opposite approach on Reddit — Womprat, a lightweight utility designed specifically for tracking small aerial targets in video. Its defining feature: it runs entirely on CPU, requires no GPU, and needs no trained models whatsoever.

The project is released under the Apache 2.0 license and hosted on GitHub (punklabs-ai/womprat). Apache 2.0 is one of the most permissive mainstream open-source licenses available — it allows free use, modification, and distribution, including for commercial purposes, and provides explicit patent grant clauses that protect users from patent litigation. In practice, this means companies can integrate Womprat directly into commercial products without legal concerns. For resource-constrained environments or developers looking to avoid complex model deployment pipelines, this is a genuinely practical lightweight alternative.

How It Works: A Return to Classical Computer Vision

Automatic Acquisition and Frame-to-Frame Tracking

Womprat's core workflow consists of two stages: target acquisition and frame-to-frame tracking.

During acquisition, the tool automatically identifies moving objects or high-contrast candidate targets in the video frame. Once a candidate is locked, the system tracks it frame by frame, continuously maintaining a fix on the target's position. The entire process relies on classical computer vision algorithms rather than the semantic understanding of a neural network.

The classical algorithms employed here come with a deep technical heritage. Motion detection is typically achieved through Frame Differencing or Background Subtraction (e.g., MOG2, KNN algorithms), which identify moving objects through pixel-level differences. Contrast analysis uses brightness and color differences between the target and background for segmentation. For tracking, classical trackers such as KCF (Kernelized Correlation Filters), MOSSE, and CSRT maintain tracking state by building appearance models of the target across consecutive frames. These algorithms have far lower computational complexity than deep neural networks, deliver strong real-time performance, and require no training data.

This design yields a significant practical advantage: an extremely low computational barrier. The tracking system runs smoothly on ordinary hardware without a GPU — including embedded and edge computing environments — dramatically lowering deployment costs and hardware dependencies.

Optional Manual Lock-On Mode

Beyond automatic acquisition, Womprat also offers a manual Lock-On mode as a supplementary option. Users can manually draw a bounding box around a specific target in the frame, directing the system to focus on that object. This is particularly useful when automatic acquisition struggles to differentiate between multiple targets, or when you need to precisely designate a target of interest.

Validated Use Cases and Known Limitations

Industry Context for Small Aerial Target Tracking

Small aerial target tracking is one of the core technical challenges in security, Counter-UAV (C-UAV), and low-altitude airspace management. As consumer drones proliferate, demand for low-altitude safety monitoring around airports and critical infrastructure has surged. Traditional radar has limited effectiveness against small, slow, low-flying targets, making visual tracking solutions increasingly attractive. The difficulty lies in the fact that targets are tiny in pixel terms (sometimes just a few pixels), backgrounds are complex (sky, trees, buildings), and target movement patterns vary widely. Womprat is tuned specifically for this vertical use case, reflecting a highly focused design philosophy.

Tuned and Validated Across Six Datasets

According to the authors, the current configuration profile has been tuned and validated across six different datasets. This means the tool isn't only effective in a single scenario — it demonstrates a degree of generalization. For tracking small aerial targets like drones and other flying objects, multi-dataset validation significantly improves reliability.

The Inherent Limitation: No Semantic Understanding

The authors are also candid about the tool's boundaries: automatic acquisition is not semantic classification.

Understanding this limitation requires grasping the fundamental difference between classical methods and deep learning. Deep learning models, trained on massive labeled datasets, learn high-dimensional mappings from pixels to semantic labels (e.g., "drone," "bird," "airplane"), enabling them to distinguish between visually similar objects of different classes. Classical computer vision methods, by contrast, can only describe low-level visual features: motion speed, outline shape, brightness contrast, and so on. They fundamentally detect anomalies rather than understand content.

In other words, Womprat can detect that "something is moving" or that "a high-contrast object is present," but it doesn't understand what that object is. As a result, the tool still faces challenges with birds and background clutter — because birds and small drones are highly similar in motion patterns and visual characteristics, classical methods have almost no algorithmic way to distinguish between them, nor can they reliably filter out interference in complex backgrounds.

This highlights the fundamental trade-off between classical computer vision and deep learning: the former wins on being lightweight and interpretable; the latter wins on semantic understanding. Both paths involve trade-offs, and neither is universally superior.

Typical Use Cases

At a time when AI tracking tools are increasingly "heavy-weight," Womprat represents a pragmatic engineering philosophy — not chasing state-of-the-art model performance, but focusing on doing tracking well within limited compute budgets.

The following scenarios are particularly well-suited for Womprat:

Edge and embedded systems: No GPU dependency means stable operation on low-power devices like Raspberry Pi and ARM development boards. In edge computing scenarios, data is processed near the source, satisfying real-time requirements while avoiding the bandwidth and privacy costs of streaming video to the cloud.
Latency-sensitive applications: Classical CV algorithms typically offer lower inference latency without waiting for neural network forward passes.
Applications requiring interpretability: The algorithm logic is grounded in well-established mathematical principles, offering high transparency that's easy to debug, audit, and trust — especially important in safety-critical domains.
Rapid prototyping: No training data or models needed; it works out of the box with minimal deployment overhead.

As an open-source project, the authors have extended a clear invitation to the community: technical feedback and challenging test footage are welcome. This open attitude helps the tool continue to improve against real-world complex scenarios — especially for hard cases like bird discrimination and clutter interference.

Summary

Womprat is a reminder that in today's deep-learning-dominated landscape, classical computer vision methods still hold irreplaceable value. For the specific problem of small aerial target tracking, a lightweight tool that requires no GPU, no pre-trained model, and is free and open-source is absolutely worth serious consideration by developers and researchers.

That said, its semantic limitations are a reminder that understanding a tool's boundaries is a prerequisite for choosing it. In scenarios that demand precise target classification (such as a counter-drone system that must distinguish birds from UAVs), complementary approaches like deep learning semantic classifiers may still be necessary. But as a lightweight, transparent, and easily deployable tracking primitive, Womprat has clearly carved out a distinct niche — it doesn't represent a technical compromise, but rather a precise engineering choice made under specific constraints.