Volleyball Real-Time Analytics
AI-Powered Real-Time Volleyball Serve Classification Through iPhone Video Streaming
problem
Our client, a volleyball head coach, sought to leverage AI technologies to enhance analytics production and adapt game strategy dynamically during matches. Recognizing human activities from video streams posed significant challenges due to factors like background clutter and changes in scale, lighting, and appearance. Moreover, the process consumes considerable processing time and incurs high costs. The client aimed to detect volleyball player actions using just one iPhone 12, making the serve classification process quick and cost-effective while automating match processing for coaches.
problem
Our client, a volleyball head coach, sought to leverage AI technologies to enhance analytics production and adapt game strategy dynamically during matches. Recognizing human activities from video streams posed significant challenges due to factors like background clutter and changes in scale, lighting, and appearance. Moreover, the process consumes considerable processing time and incurs high costs. The client aimed to detect volleyball player actions using just one iPhone 12, making the serve classification process quick and cost-effective while automating match processing for coaches.
solution
The project aimed to develop a machine learning (ML) model capable of classifying volleyball serves from video streams with an F-1 score of 0.85. Key requirements included optimizing the system for near real-time performance and integrating the model into iOS 12. We utilized TensorFlow to build the model and CoreMLTools Python Lib to convert it into a format suitable for CoreML. Integration into iOS 12 was accomplished using the CoreML framework and Xcode IDE, ensuring seamless deployment on iPhone 12 devices
solution
The project aimed to develop a machine learning (ML) model capable of classifying volleyball serves from video streams with an F-1 score of 0.85. Key requirements included optimizing the system for near real-time performance and integrating the model into iOS 12. We utilized TensorFlow to build the model and CoreMLTools Python Lib to convert it into a format suitable for CoreML. Integration into iOS 12 was accomplished using the CoreML framework and Xcode IDE, ensuring seamless deployment on iPhone 12 devices
results
Initially, our team explored using the ActionClassifier from createML but encountered limitations when dealing with multi-person action classification. Subsequently, they experimented with the ImageClassifier tool, focusing on event classification instead of action classification. Despite facing challenges with accuracy and training time, the team persevered, eventually preparing a dataset comprising sequential grayscale frames for model training. By the project's end, the team delivered a serve detection model for volleyball games operating at real-time speed (26 FPS) with an F1 score of 0.87, achieved within a six-week timeframe.
results
Initially, our team explored using the ActionClassifier from createML but encountered limitations when dealing with multi-person action classification. Subsequently, they experimented with the ImageClassifier tool, focusing on event classification instead of action classification. Despite facing challenges with accuracy and training time, the team persevered, eventually preparing a dataset comprising sequential grayscale frames for model training. By the project's end, the team delivered a serve detection model for volleyball games operating at real-time speed (26 FPS) with an F1 score of 0.87, achieved within a six-week timeframe.
tech stack
tech stack
