OpenCV Detection of Athletes in Long Jumping Videos

Ailing Song, Kai ChenProceedings of the 2013 International Conference on Information, Business and Education Technology (ICIBET 2013) 2013 · DOI

Abstract The precise analysis of long jumper technical videos provides critical data for coaches to optimize training programs. To enable automated analysis, athletes must first be accurately detected and tracked. This paper explores an athlete detection method using the OpenCV library, specifically focusing on an improved background subtraction technique designed to handle the dynamic lighting conditions typical of outdoor sports environments.

What Problem Does This Paper Address?

In competitive long jumping, success is determined by the seamless integration of the run-up, take-off, flight, and landing phases. While video analysis is essential for coaching, automating the extraction of technical data is challenging.

Standard background subtraction often relies on the Surendra algorithm, which typically uses a fixed brightness threshold. In real-world scenarios—such as outdoor stadiums with shifting shadows and changing sunlight—a static threshold fails. It may either misidentify light patches as moving objects or fail to capture the athlete's full contour. This research proposes a "self-adapting" detection process to overcome these environmental limitations.

Methods

The proposed solution is built on the OpenCV framework and utilizes an "Improved Surendra Algorithm." The workflow is designed to be robust against noise and lighting shifts through a multi-stage process:

  1. Preprocessing: Video frames are cleaned using Median and Gaussian filters to minimize sensor noise.
  2. Dynamic Thresholding (OTSU Integration): Instead of a hardcoded threshold, the system employs the OTSU algorithm. This automatically calculates an optimal threshold for every frame based on the image's histogram, allowing the system to adapt to lighting changes in real-time.
  3. Background Updating: The system maintains an iterative background model ($B_k$). It uses an inter-frame difference method to create a mask, ensuring the background model is updated without incorporating the moving athlete.
  4. Foreground Extraction: The moving target is isolated by calculating the difference between the current frame and the updated background.
  5. Morphological Refinement: Finally, dilation and erosion operations are applied to the foreground mask to fill holes and remove "salt-and-pepper" noise, resulting in a solid contour of the athlete.
Flowchart of the improved background subtraction algorithm.

Datasets

The evaluation was performed on custom-recorded "technological videos" of long jumpers. These videos serve as a practical testbed for demonstrating the algorithm's ability to maintain detection accuracy across frame sequences in a coaching environment, rather than relying on standard public benchmarks.

Key Insights and Ideas

  • Adaptability is Essential: The primary innovation is the transition from static to dynamic thresholding via OTSU. This shift makes background subtraction robust enough for practical, outdoor use.
  • Hybrid Logic: The method effectively combines background subtraction (for global context) with inter-frame difference (for local motion masking). This prevents "ghosting" effects where a moving object's previous position is incorrectly identified as part of the background.
  • Contour-Focused Design: Unlike simple bounding-box detectors, this method prioritizes extracting a complete contour. For a coach, the exact angle of a limb or the curvature of the spine during the flight phase is more valuable than a simple coordinate.

Key Results

Qualitative analysis of the video sequences demonstrated the algorithm's effectiveness:

  • Threshold Evolution: The system successfully adapted its internal parameters as lighting changed. In one test case, the threshold evolved from 60 in the initial frame to 110 by the 202nd frame.
  • Contour Integrity: Despite significant shifts in environmental lighting, the system maintained a complete and accurate description of the athlete’s position, size, and shape.
  • Environmental Stability: The self-adapting background model remained stable and was largely unaffected by external environmental factors.
Athlete segmentation results showing the original frame, background difference, and final thresholded foreground.

Limitations and Future Work

While the detection logic is sound, several areas remain for future development:

  • Quantitative Benchmarking: Future studies should include quantitative metrics (e.g., Precision, Recall, F1-score) against standard datasets to compare performance with modern deep learning models like YOLO or Mask R-CNN.
  • Computational Performance: The paper does not specify the frames-per-second (FPS) achieved. Real-time coaching applications require high-speed processing to be viable during live training sessions.
  • Multi-Target Tracking: The current work focuses on detection. The next step is implementing robust tracking to maintain the athlete's identity if officials or other participants enter the field of view.

Summary updated

Abstract

The precise analysis of long jumper's technical videos can provide the important information for the coaches to formulate the training program. To realize the analysis of technical videos, we must firstly detect and track the long jumpers in the video. OpenCV technique (open source computer vision library supported by the Intel company) is preferable in the field of image processing. This paper aims to study how to detect the athletes in long jumper's technical video based on OpenCV technique and provide some technical reference to the related research.
BibTeX
@inproceedings{song2013opencv,
  author = {Ailing Song and Kai Chen},
  title = {{OpenCV Detection of Athletes in Long Jumping Videos}},
  booktitle = {Proceedings of the 2013 International Conference on Information, Business and Education Technology (ICIBET 2013)},
  year = {2013},
  doi = {10.2991/icibet.2013.3},
}
Figure