Alif, Bosch Sensortec, and Edge Impulse have simplified the creation of edge sensing applications that classify complex motions and identify gestures.
Alif Semiconductor is collaborating with Bosch Sensortec and Edge Impulse to combine hardware and software in reference designs that remove the complexity of creating precision motion-sensing products, while drastically increasing performance, lowering power consumption, and establishing airtight security. Applications that benefit include game controllers, predictive maintenance, and monitoring status of industrial machinery and cargo.
Typically, creating and training an artificial intelligence/machine learning (AI/ML) framework that consumes dynamic sensor input, then deploying the model to a hardware platform and optimizing the model on that specific hardware for speed, accuracy, and power efficiency is a very difficult task for developers. Alif and its partners have completely removed this complexity by offering reference designs based on the secure, low-power E3 MCU from Alif’s Ensemble family, Bosch Sensortec’s BMI323, an affordable Inertial Measurement Unit (IMU) with high accuracy and power efficiency, and Edge Impulse’s development platform for machine learning.
The E3 MCU employs dual processing domains that uniquely accelerate machine learning. The E3’s High Efficiency processing domain is meant to be always on to sense the environment, while the E3’s High Performance processing domain wakes as needed to rapidly execute heavy workloads and returns to sleep. In the reference design the BMI323 IMU feeds precision acceleration and angular rate data to the MCU. Edge Impulse’s platform is used to rapidly train an ML model to identify complex continuous gestures such as multiple characters, symbols, and anomalies in repeating complex movement patterns.
The results are astonishing.
In less than one hour, a ML model can be created, trained, and deployed to the E3 MCU to detect and identify a complex multi-directional motion pattern and translate it to a symbol. While running, the time to complete the inference operation required to identify the gesture pattern is only 280μs.
By comparison, a traditional model trained to identify movement in a single direction deployed on a traditional Cortex-M7 MCU required 2.3ms for the inference operation to complete, and a traditional Cortex-M4F MCU required 12 msec. These results are nine times slower and 43 times slower than Alif’s results, and the simplistic nature of the detectable motion is very limiting for real world use-cases
“This is truly an amazing uplift in performance and potential,” said Stefan Finkbeiner, CEO at Bosch Sensortec. “The increase in inference speed directly translates to lower system-level power consumption and could also be used for this kind of solution to operate at significantly higher symbol rates. This opens the door to deploying smart sensors in ways that are not achievable today.”
“Edge Impulse is specifically designed to leverage the power of advanced tools like Alif’s processors and Bosch Sensortec’s sensors, enabling groundbreaking development without the complexity and lead time traditionally needed for ML applications,” said Zach Shelby, CEO and Co-founder of Edge Impulse. “The performance demonstrated with this combination is going to be very attractive to enterprises looking to build the next generation of technology solutions.”
“One of the key drivers behind the architecture of the Ensemble family has been to deliver a higher degree of integrated high-level functions to edge platforms than ever seen before,” said Reza Kazerounian, President and Co-Founder at Alif Semiconductor. “Being able to handle this performance level of machine learning workloads within the power and price budget of an MCU is unmatched in the industry today.”