Javelin Throw Analysis System with YOLO11: OBB Detection + Pose Estimation in Action
A developer built a full javelin throw analysis pipeline from broadcast video using YOLO11's multi-model architecture.
A developer built a complete javelin throw analysis system on YOLO11, working directly from broadcast footage without specialized equipment. The system detects javelins, estimates athlete pose, calculates joint angles, identifies throwing phases, and plots flight trajectories. It uses a three-layer architecture — OBB detection, YOLO11 Pose, and multi-object tracking with ByteTrack and BoT-SORT. Key engineering challenges include motion blur, false detections, ID switches, and camera movement. The project is open-source and offers a reusable reference architecture for sports video analysis.
When Computer Vision Meets Track and Field
Sports performance analysis is emerging as a major application area for computer vision. Recently, a developer shared on Reddit a javelin throw analysis system built on YOLO11, sparking widespread discussion. The system can automatically detect javelins from broadcast footage, track athlete movements, calculate joint angles, identify throwing phases, and plot the javelin's flight trajectory after release.
Unlike traditional sports analysis that relies on specialized equipment and manual annotation, this system works entirely from standard broadcast video — demonstrating the impressive real-world capabilities of modern object detection and pose estimation technology.
Tech Stack Breakdown: Multi-Model Collaboration
This system doesn't rely on a single model. Instead, it orchestrates multiple components to handle the complexity of sports video analysis. The developer integrated a set of leading computer vision tools into a complete end-to-end pipeline.
Core Components
The tech stack includes the following key elements:
- YOLO11s-OBB (Oriented Bounding Box): A custom-trained model dedicated to javelin detection. Since a javelin is a long, slender object, traditional axis-aligned bounding boxes (HBB) struggle to fit it precisely. OBB rotates to any angle, conforming tightly to the javelin's actual shape and significantly improving detection accuracy.
- YOLO11 Pose: Used for human pose estimation — detecting skeletal keypoints to locate the throwing arm and compute joint angles.
- OpenCV: Handles image processing, trajectory drawing, and angle visualization.
- ByteTrack and BoT-SORT: Two multi-object tracking algorithms that maintain stable tracking of both the javelin and athlete across consecutive frames, mitigating ID-switch issues.
This "detection + pose + tracking" combination represents the typical architectural approach in modern sports video analysis.
Core Capabilities of the System
Based on the developer's description, the javelin analysis system covers a broad and complete set of analytical dimensions relevant to the sport.
From Object Detection to Action Understanding
The system first detects and tracks the javelin using the OBB model, while simultaneously using pose estimation to identify the athlete's throwing arm. Building on this, it calculates smoothed elbow angle — a particularly important metric in sports technique analysis, as elbow joint angle is one of the key indicators for evaluating throw quality.
Going further, the system can also identify distinct throwing phases (e.g., run-up, crossover steps, final effort, release) and plot the flight trajectory after the javelin leaves the hand. This phase recognition and trajectory visualization gives coaches and athletes an intuitive, data-driven basis for technical feedback.
Technical Challenges in Real-World Scenarios
The developer was candid about the many difficulties encountered during development. These challenges closely mirror the universal pain points of applying computer vision to real-world video.
The Gap Between Lab Conditions and Broadcast Footage
Key technical challenges include:
- Motion Blur: A javelin traveling at high speed creates blur in the frame, making detection difficult.
- False Detections: Other elongated objects in the background can be mistaken for a javelin.
- Tracking ID Switches: When a target briefly disappears or becomes occluded and then reappears, the tracker may assign a new ID, causing trajectory fragmentation.
- Camera Movement: Broadcast footage pans and zooms to follow athletes, introducing significant instability for tracking.
- Missed Detections: In certain frames, the javelin or athlete may not be detected at all.
- Selecting the Right Athlete: Broadcast scenes often contain multiple people — referees, other competitors, spectators — and the system must accurately lock onto the thrower.
These issues may not surface on controlled benchmark datasets, but they are routine obstacles in real television broadcast material. The developer's willingness to confront and work through each of these is precisely what gives this project its practical value.
Takeaways and Practical Advice for Developers
This project serves as an excellent reference for developers looking to enter sports video analysis or similar computer vision applications. The full source code is open-sourced on GitHub (github.com/chahmadraza89/computer_vision) for learning and further development.
Key Technical Insights Worth Borrowing
First, for elongated or irregular-shaped objects, OBB is a better fit than traditional HBB — a nuanced but impactful model selection detail that's easy to overlook. Second, multi-model collaboration outperforms a single model — having dedicated models for detection, pose estimation, and tracking allows the system to handle far more complex real-world tasks. Finally, choosing and tuning the right tracking algorithm is critical — the combination of ByteTrack and BoT-SORT reflects a deliberate engineering tradeoff in pursuit of tracking stability.
As the YOLO model family continues to evolve with improving detection accuracy and inference speed, real-time or near-real-time sports analysis applications are becoming increasingly feasible. It's foreseeable that computer vision-based analytical tools will play an ever-growing role in professional athletic training, broadcast enhancement, and amateur sports coaching alike.
Related articles
Catalyst: A Vision for an Enzyme-Like Testing Framework for AI Agents
A developer shared Catalyst on Reddit, an Enzyme-inspired framework for AI Agents, exploring why agents need observable, testable dev tools and the design philosophy behind them.
The Real Capability of AI Coding Agents: Best Models Complete Only 35% of Feature Development Tasks
The 'Agents on Rails' benchmark finds top AI models complete only 35% of feature development tasks. What this means for coding agents and developer teams.
How to Prevent Duplicate Refunds After an AI Agent Crashes: CellaFlow's Durable Execution Approach
How can AI agents avoid duplicate refunds after a crash without deadlocking workflows? CellaFlow uses durable execution, shared work identity, leases, and fencing to solve safety and liveness in multi-agent systems.