A Low-Cost Approach for Sports Performance Assessment Using Computer Vision and Artificial Intelligence Tools

1Department of Computer Science, University of Puerto Rico at Río Piedras 2Facultad de Ciencias en la Salud, Universidad Manuela Beltrán, Bogotá, Colombia 3Department of Physical Education and Recreation, University of Puerto Rico at Río Piedras Introduction A Low-Cost Approach for Sports Performance Assessment Using Computer Vision and Artificial Intelligence Tools You Only Look Once: Yolo 2 References The system developed takes sports training videos and runs the Yolov11 algorithm over them, prompting the user to highlight key areas of the video (such as goals and ball starting positions) which it later uses alongside the Yolo model's output to analyze the training and evaluate player performance. Returns the attempt's direction, accuracy, as well as the player's reaction time. The program also outputs the analyzed video (displaying ball trajectory and size which has to be predicted when not detected by Yolo) and various graphs/data for further analysis, (See Figure 3). Currently we use Google Colab [5] to run the program, as it offers GPUs better suited for computationally intensive tasks. Results The training environment that inspired this project, developed by Javier Osorio and based on studies by Musculus et al. [1] and Knöbel & Lautenbach [2], plays an important role in the functioning of the software. We are developing an ecologically valid environment, that is, one that integrates perceptual, cognitive, and motor skills within specific sports contexts [3]. The system consists of a projector to generate a visual stimuli for the player and a smartphone to capture videos containing the athlete’s reactions. Figure 1 shows a schematic representation of the training environment, in which the player is presented with two possible soccer field goals. Figure 2 illustrates an example of a player’s kick after the video has been processed using YOLO, with ball-tracking added during post-processing. Traditional sports training often relies on manual observations and subjective feedback, which can be slow, inconsistent, and prone to error. Existing sports training systems and specialized software tend to be expensive and difficult to implement, making them inaccessible for many university sports teams. In this work, we explore how artificial intelligence and computer vision can be applied to track, analyze, and evaluate athletic performance more objectively and efficiently within a low-cost training framework. Training Environment 1. Musculus, L., Lautenbach, F., Knöbel, S., Reinhard, M. L., Weigel, P., Gatzmaga, N., ... & Pelka, M. (2022). An assist for cognitive diagnostics in soccer: two valid tasks measuring inhibition and cognitive flexibility in a soccer-specific setting with a soccer-specific motor response. Frontiers in psychology,13, 867849. 2. Knöbel, S., & Lautenbach, F. (2023). An assist for cognitive diagnostics in soccer (Part II): Development and validation of a task to measure working memory in a soccer-specific setting. Frontiers in psychology,13, 1026017. 3. Renshaw, I., Davids, K., Araújo, D., Lucas, A., Roberts, W. M., Newcombe, D. J., & Franks, B. (2019). Evaluating weaknesses of “perceptual-cognitive training” and “brain training” methods in sport: An ecological dynamics critique. Frontiers in psychology,9, 2468. 4. Ultralytics. (n.d.). YOLO thread-safe inference. Ultralytics YOLO Docs. Retrieved August 19, 2025, from https://docs.ultralytics.com/guides/yolo-thread-safe-inference/ 5. Google. (n.d.). Google Colaboratory. Retrieved August 19, 2025, from https://colab.research.google.com Our system uses YOLO or "You Only Look Once," Ultralytics' real time object detection deep learning model, which through its robust algorithm is able to quickly and accurately identify different objects in images [4]. Yolov8. We originally started with some demos using Yolov8, a version of the model that is able to process images very quickly but sacrifices some accuracy. Yolov11. We switched to Yolov11 after some time due to Yolov8's issues with detecting objects consistently and accurately at longer ranges. We decided to use Yolov11l, a version of Yolov11 which has around 25.3 million parameters (making the program slower, but much more accurate). 2 Acknowledgements This work was partially supported by the Natural Sciences Dean’s Office at UPR Río Piedras,Dr. Alvarez’s Seed Funds, and Dr. Orozco’s FADI Funds. Special thanks to Javier Osorio for generously sharing ideas from his doctoral dissertation. Gabriel Torres1Carlos Vázquez1Javier Osorio2,3 Edusmildo Orozco1Michael Alvarez1,* [email protected] [email protected] [email protected] [email protected] [email protected] LATIN AMERICA HIGH PERFORMANCE COMPUTING CONFERENCE September 22-26 Kingston, Jamaica Conclusions This system uses the YOLOv11 object detection model to analyze sports training videos by automatically identifying key elements like goal areas and ball positions. It computes metrics such as shot accuracy, reaction time, and direction, and outputs annotated videos with predictive ball trajectories alongside performance graphs. Running on Google Colab with GPU acceleration, the implementation supports parallel processing through Python multiproccesing, offering significant speed gains. Designed for realistic game scenarios, affordable cost, and a simple implementation, the tool shows promising potential to enhance coaching feedback and support player development in university-level sports programs. Additionally, the integration of gamification strategies such as point systems, challenges, and interactive feedback could further increase player engagement, motivation, and learning outcomes, making the tool even more effective in training environments. Figure 1. Experimental setup: projection-based environment displaying two possible soccer fields for perceptual and decision-making training. Figure 2. Example of the output video processed with YOLO and trajectory highlighted in post-processing. Figure 3. Trajectories obtained after post-processing YOLO labels to quantify the player’s accuracy. The dotted line indicates the point of the shots, and the soccer fields are included as part of the player’s visual context.