Case study
AiSport company is building an Ai Coach, who helps to improve the technique of the trainings via real-time personal feedback. To reach this goal their team have developed ML models for sports training and analysis. They provide real-time data processing with a high degree of accuracy. ThinkAiAgency was hired to work with AiSport's team in order to improve their technology.
The main challenge was to improve the accuracy and efficiency of their existing models for detecting human body joints, motion pattern extraction, mistakes recognition, pattern matching, repetition counting, and instant feedback. In order to achieve this, the team needed to experiment with different approaches and techniques.
Another challenge was to make the technology easily accessible and scalable to individual clients and enterprises worldwide. ThinkAiAgency was tasked with providing access to AiSport's proprietary technology using their API and custom-built integration.
We worked with AiSport's ML engineers and experts to experiment with both top-down and bottom-up approaches for detecting human body joints. The bottom-up approach was found to produce better performance compared to the top-down approach. They also leveraged MobileNetV2 for feature extraction, and used CenterNet prediction heads to obtain density maps of the probability distribution of the target joints. The prediction heads were then filtered with non-max-suppression to generate human body skeletons. The performance of the models was estimated using various metrics such as MAE, PA-MPJPE, Geodetic distances, DTW, etc.
ThinkAiAgency also helped AiSport to make their technology easily accessible and scalable to individual clients and enterprises worldwide. This allowed enterprises to get started quickly with 2D and 3D skeleton recognition & analysis, motion patterns matching, and equipment tracking as an out-of-the-box API that can be used by any application or service to receive accurate processing of images or videos.