Normative Evaluation Method of Long Jump Action Based on Human Pose Estimation

Xiugang Gong, Xinyuan Geng, Guangjun Nie, Tao Wang, Jiajun Zhang, Jiachuan YouIEEE Access 2023 · DOI

TL;DR

This paper introduces an objective, AI-driven system for evaluating standing long jump technique using lightweight human pose estimation. By combining a MoveNet-based model with Dynamic Time Warping (DTW) and cosine similarity, the system achieves 91.3% accuracy in movement analysis and provides real-time corrective feedback on low-power edge devices like the Raspberry Pi.

What Problem Does This Paper Address?

In traditional sports training, evaluating a student's form is often a subjective process. Coaches rely on their "trained eye" to spot errors, which can lead to inconsistent feedback, slow progress, or even sports-related injuries due to improper technique.

While wearable sensors (like accelerometers) offer objective data, they are cumbersome to put on for every student in a large class. Computer vision offers a non-invasive alternative, but standard models are often too computationally heavy for real-time use in a school gym or are easily fooled by variations in a student's height, weight, and jumping speed. The authors sought to bridge this gap by creating a system that is both mathematically rigorous and lightweight enough to run on embedded hardware.

Methods

The researchers developed a multi-stage pipeline to move from raw video to a "score" and corrective suggestions:

  1. Lightweight Pose Estimation: They utilized the MoveNet algorithm with a MobileNet-v3 backbone. This architecture is designed for speed, using a feature pyramid network (FPN) to generate heatmaps for the body's center and 17 key joints.
  2. Key Action Extraction: A long jump consists of hundreds of video frames, but only a few are critical. The system automatically identifies four phases—take-off, aerial, landing, and buffering—using Dynamic Time Warping (DTW). DTW is crucial here because it allows the system to compare a student's jump to a "standard" jump even if they move at different speeds.
  3. Normative Analysis: To score the jump, the system uses Cosine Similarity to compare joint angles. Unlike Euclidean distance (which can be distorted by a student's height), cosine similarity focuses on the direction and angle of the limbs.
  4. Trajectory Filtering: The system applies a filtering process to the foot's trajectory and uses a Riemann integral to calculate the deviation from a standard parabolic arc.

Datasets

The authors used a hybrid dataset to train and validate their model:

  • COCO (2017) Dataset: 1,362 samples were curated from this standard benchmark to provide a foundation for general human pose estimation.
  • Proprietary Long Jump Dataset: The team collected and labeled additional samples specifically of standing long jumps to fine-tune the model for the unique postures found in the sport.
  • Data Augmentation: Techniques like mirroring were used to increase the diversity of the training set and minimize "misplaced keypoints" during fast movements.

Key Insights and Ideas

  • Overcoming Physical Diversity: The most significant "aha moment" is the move away from coordinate-based distance. By using cosine similarity for joint angles and DTW for temporal alignment, the system becomes "body-type agnostic." It doesn't matter if the student is 5'2" or 6'2"; the normativity of the angle remains the same.
  • The "Ladder Similarity" Approach: Instead of just giving a pass/fail grade, the system uses a weighted scoring formula. It prioritizes certain joints over others depending on the phase of the jump (e.g., leg angles are more critical during the take-off than arm positions).
  • Edge Computing Viability: The choice of MobileNet-v3 and specific downsampling techniques proves that sophisticated sports analysis doesn't require a server rack; it can happen on a $50 micro-computer at the edge of the track.

Key Results

  • High Accuracy: The proposed method achieved an average accuracy of 91.3% across the four key actions.
  • Superior to Human Subjectivity: In a head-to-head comparison, the system's accuracy (91.3%) significantly outperformed the subjective judgments of three physical education teachers, whose average accuracy rate was only 81.0%.
  • Performance on Hardware: On a Raspberry Pi 4B (1.8GHz), the model maintained a frame rate of 10 FPS. While this sounds low compared to gaming, it is sufficient for analyzing the key phases of a jump and is 10x faster than more complex models like AlphaPose or LightWeight OpenPose on the same hardware.
  • Actionable Feedback: The system successfully generated specific text-based suggestions, such as "The upper body is too high. Lower is recommended," which closely matched expert coaching advice.

Limitations and Future Work

While the results are promising, the paper leaves room for expansion:

  • Single-Plane Analysis: The current model appears to rely on a side-view (sagittal plane) for the best results. Future iterations could incorporate multi-camera setups for 3D pose estimation.
  • Specific Sport Focus: The system is currently tuned for the standing long jump. Adapting it to the running long jump or other complex movements like high jumps would require new "standard" action sequences and potentially higher frame rates to capture faster approach speeds.
  • Environmental Factors: The paper does not deeply explore how the model performs in varied lighting conditions or cluttered backgrounds (like a busy school gym), which are common in real-world deployments.

Summary updated

Abstract

Addressing the issue of the lack of objective quantitative evaluation in training long jump events, this study presents a normative analysis method based on human pose estimation and similarity measures. By training a lightweight human pose estimation model, this method can run on low-delay embedded devices. In line with key movements, the proposed method designs a normative analysis for long jump actions, which yields a measurement of the movements’ adherence to the standard and provides corrective suggestions. Experimental results indicate that the accuracy of this approach in analyzing the standardization of long jump action reaches 91.3%. As a result, it holds significant application value in various scenarios, including students’ long jump training and correction of long jump techniques. Furthermore, it can be extended to other practical applications beyond sports.
BibTeX
@inproceedings{gong2023normative,
  author = {Xiugang Gong and Xinyuan Geng and Guangjun Nie and Tao Wang and Jiajun Zhang and Jiachuan You},
  title = {{Normative Evaluation Method of Long Jump Action Based on Human Pose Estimation}},
  booktitle = {IEEE Access},
  year = {2023},
  doi = {10.1109/access.2023.3330757},
}
Figure