Project closure

Neuromorphic AI enables precise and energy-efficient action recognition in real time

In sports, especially in tennis, the precise analysis of movements is becoming increasingly important. With the completion of the SpikingBody research project, fortiss has developed an innovative solution that uses neuromorphic AI to recognize player actions in real time and with exceptional energy efficiency. The system combines event-based image sensors with specialized neuromorphic hardware to enable precise, low-latency, and energy-efficient capture of human movements.

With the SpikingBody project, fortiss is further expanding its leading expertise in biologically inspired AI technologies in the field of neuromorphic computing. Under the leadership of Michael Neumeier, the team of researchers explored the use of neuromorphic algorithms based on spiking neural networks (SNNs) and Intel's neuromorphic research chip Loihi for real-time action recognition. The research, funded by the Bavarian Research Foundation, opens up promising prospects for applications in human-machine interfaces, particularly in the fields of robotics, industrial automation, and interactive assistance systems.
 

Neuromorphic intelligence for intuitive human-machine interaction

Understanding human actions is fundamental to the safety, acceptance, and interactivity of robots and machines. However, effective recognition through human-machine interfaces requires high detection accuracy, fast response times, and low energy consumption. Conventional systems often reach their limits in this regard. Neuromorphic technologies, which are modeled on the architecture and functioning of the human brain, make it possible to optimize action recognition systems according to these criteria.

SpikingBody relies on a novel approach: combining event-based vision sensors with SNNs implemented on neuromorphic hardware. The event-based image sensors respond only to changes in the environment, rather than capturing complete images at high data rates. This reduces computational load and provides high temporal resolution with minimal latency. The captured sensor data is processed directly as spikes in an SNN, which runs energy-efficiently on Intel’s research platform Loihi. The network recognizes and interprets actions in real time – with the ability to learn new actions 'on the fly' without needing to retrain the entire system.
 

Neuromorphic AI is transforming motion analysis in tennis

This technology has been specifically applied to tennis action recognition and offers new possibilities for analyzing and responding to player movements in real time. Thanks to the neuromorphic approach, detection latency can be reduced to the point where the type of movement is recognized in real time. The event-driven camera, which generates a very compact image stream with virtually no latency, makes it possible to capture very fast movements without motion blur, thereby enabling reliable speed estimations.

The system developed by fortiss uses just a single event camera to recognize the type of a tennis player’s movement (such as serve, forehand, or backhand) in real time. Based on this recognition, initial conclusions can be drawn or more expensive, high-precision systems can be activated to enable more detailed analysis – for example, for broadcasting matches on television. At the same time, it offers sports clubs a more affordable and user-friendly alternative to professional analysis and training systems.

Neuromorphic Computing
Output of the event camera during the forehand stroke (tennis).
Neuromorphic Computing
Pose estimation on the forehand stroke (tennis).

Two-stage learning as the key to efficiency

A key result of the SpikingBody project is the development of a two-stage learning approach, consisting of offline pretraining and online learning in real time. This allows the system to recognize new actions during operation and continuously adapt – a crucial advantage for human-machine interaction, robotics, and industrial automation. In the case of the tennis application, this means that the system can adapt to new players or even learn new actions.

The event processing network performs two tasks simultaneously: classifying the actions and estimating the body position by reconstructing the joint positions of the player. This not only enables the categorization of the action (serve, backhand, etc.) but also the estimation of the angles and speeds of the shots.
 

Advances in the development of efficient, adaptive AI technologies

fortiss sets new standards in energy-efficient and adaptive AI research with SpikingBody. The developed methods and prototypes open up innovative application fields in industrial control systems, robotics, and smart user interfaces. From contactless machine controls to interactive assistance systems and novel control solutions for robotic systems – SpikingBody demonstrates the potential of powerful, learning-capable, and resource-conserving AI technologies.

An excellent example of the versatility of this technology is a sound art installation developed by Dr. Axel von Arnim, Head of the Neuromorphic Computing competence field, and his team. Through real-time processing, the system responds to the movements of visitors and creates interactive music spaces that dynamically change. The combination of technical innovation and artistic visualization gives the impression of a close, creative connection between humans and machines. This installation was presented at the Festival of the Future in Munich in July 2024.

Building on these results, fortiss, together with partners from industry and science, will develop the next generation of smart, adaptive human-machine interfaces. With SpikingBody, fortiss once again makes a statement for the future of AI and robotics – efficient, learning-capable, and energy-optimized.

  Marketing & press

Your contact

Marketing & press

presse@fortiss.org