- The paper introduces PhysTrackX, a Python-based laboratory framework that combines NanoTrack machine-learning video tracking, Tesseract OCR, geometric measurement, calibration, plotting, and CSV export in one workflow.
- The demonstrations show how synchronized trajectories and sensor readings can analyze damped rolling motion and balloon pressure–radius behavior, including phase-space interpretation and Mooney–Rivlin model fitting.
- The software targets accessible, GPU-free physics education, but its claims of sub-pixel accuracy and robustness require quantitative benchmarks under blur, occlusion, illumination changes, and out-of-plane motion.
PhysTrackX is an open-source video analysis framework for undergraduate physics laboratories, developed at Lahore University of Management Sciences (2608.13977). It is positioned as a successor to the widely used Open Source Physics Tracker and the MATLAB-based PhysTrack, addressing two persistent shortcomings of existing tools: tracking degradation under motion blur, rapid acceleration, illumination changes, or occlusion; and the absence of native mechanisms for aligning visual kinematics with independent digital sensor readouts. The software packages machine-learning-based object tracking, optical character recognition (OCR) of in-frame digital displays, interactive geometry measurement, and automated kinematic plotting into a single streamlined graphical workflow.
Architecture and core features
The application is written in Python with a customtkinter-based GUI, OpenCV and FFmpeg for video processing, and a modular toolset that follows a natural pipeline: video loading and trimming, axis and scale calibration, tracking, plotting, and export. Pre-processing filters—Canny edge detection, Gaussian blur, median and bilateral filtering, and contrast/brightness adjustment—can be applied before tracking to stabilize performance under poor laboratory lighting or blur.
The central technical contribution is the replacement of manual point marking and classical template matching with NanoTrack, a lightweight anchor-free Siamese network executed through the ONNX runtime on CPU. The authors claim sub-pixel, drift-resistant multi-target tracking without GPU acceleration, which matters for resource-constrained teaching laboratories. A second contribution is an OCR plugin built on Tesseract: students draw a bounding box over any digital display visible in the frame (a multimeter, pressure gauge), and extracted numerical readings are appended to the trajectory with exact frame timestamps. This directly fuses multimodal data streams from a single video source, eliminating manual synchronization. Additional tools include a geometry plugin that overlays triangles and line segments, computes physical lengths from the user-defined scale factor, and derives interior angles via the law of cosines—enabling in-situ angular measurements without protractors—and a data management layer exporting coordinates, OCR values, and timestamps to CSV.
Demonstrations
Two case studies illustrate the software. In "Galileo's damped oscillator," cylinders of varying composition (solid aluminum versus hollow shells filled with fluids of different viscosities) roll down a V-shaped ramp while recorded by a smartphone camera. PhysTrackX extracts position-time trajectories and velocity-position phase-space plots; the inward spirals in phase space provide a quantitative signature of shear-induced energy dissipation across the four fill compositions.
In the "two-balloon experiment," interconnected balloons of different initial radii exchange air, and the pressure–radius relationship probes rubber hyperelasticity. PhysTrackX concurrently tracks the balloon boundaries and reads live pressure values from an in-frame digital display via the OCR module. The synchronized data capture the characteristic pressure peak, intermediate softening, and strain-hardening regime, and are fit to the Mooney–Rivlin hyperelastic model—demonstrating that quantitative constitutive-model fitting can be performed entirely from a single smartphone video.
Limitations and open questions
The paper concedes several constraints implicitly rather than explicitly. All validation is qualitative: no benchmark against Tracker or Kinovea quantifies tracking accuracy, drift, or runtime, so the claimed sub-pixel precision remains unevaluated. The OCR module depends on legible digital displays within the frame and on Tesseract's robustness to lighting and perspective distortion, neither of which is stress-tested. The case studies involve slow, planar, well-illuminated motion; behavior under occlusion, out-of-plane rotation, or severe blur—the very conditions motivating the neural tracker—is not demonstrated. Distribution is limited to Windows binaries under a BSL license, leaving cross-platform availability open.
Conclusion
PhysTrackX consolidates ML-based tracking, OCR-based sensor synchronization, geometric calibration, and plotting into an accessible open-source package aimed at reducing software-related cognitive load in instructional laboratories. Its principal claims—CPU-only sub-pixel tracking and seamless multimodal synchronization—are plausible given the underlying components but await systematic accuracy benchmarks and evaluation in more demanding experimental conditions.