From September 22 to 27, the 27th European Microwave Week (EuMW 2024) was successfully held in Paris. As a renowned event in Europe for wireless, microwave and integrated circuits, EuMW 2024 featured three major academic conferences: the European Microwave Conference, the European Microwave Integrated Circuits Conference and the European Radar Conference (EuRAD), accepting over 500 paper submissions. Calterah's paper titled "In-Cabin Detection, Localization and Classification Based on mmWave Radar with TinyML" was selected by EuRAD of EuMW 2024, as one of the only two papers from China among more than 100 accepted papers. Radar System Expert at Calterah, Dr. Zhifei Wan, was also invited to attend EuRAD 2024 to share in-depth insights on Calterah's latest in-cabin detection radar SoC technology.

In-cabin occupant detection is a crucial technology for ensuring vehicle safety, widely used in scenarios such as seatbelt reminders, airbag deployment optimization, and Child Presence Detection (CPD). The European New Car Assessment Programme (Euro NCAP) mandates that, starting from 2025, new cars must be equipped with active CPD systems to receive bonus points. Similarly, the 2024 edition of the China New Car Assessment Programme (C-NCAP) officially includes CPD as a feature for scoring additional points. Due to its advantages in privacy protection, insensitivity to lighting conditions and high detection accuracy, mmWave radar is an ideal choice for CPD applications.

However, in real-world CPD applications, in-cabin radar systems must possess anti-interference capabilities to avoid false alarms. More advanced applications require not only the ability of detecting occupant presence but also the functions to localize and classify them, which presents challenges for in-cabin detection solutions available in the market. Dr. Zhifei Wang noted, "Current radar solutions based on the Capon algorithm* have high computational complexity and may produce false alarms due to interference in complex environments. Some academic studies have attempted to use deep learning for in-cabin detection, but these models rarely consider the memory and time constraints of radar SoCs, making it difficult to deploy these models on embedded radar systems."

To tackle these challenges, Calterah has developed a real-time solution for in-cabin occupant detection, localization and classification. Calterah’s in-cabin radar solution is based on its 60 GHz mmWave radar SoC and utilizes the Tiny Machine Learning (TinyML) technology, achieving over 96 percent accuracy in occupant detection and over 90 percent accuracy in localization and classification. The solution also minimizes interference, reducing false alarms caused by movements like the shaking of vehicles or objects such as water bottles. These exceptional detection capabilities are made possible by several innovative technologies used in Calterah’s in-cabin radar solution:

1. Novel Signal Processing Flow for Enhancing Radar Robustness

Traditional radar solutions typically use power distribution in the range-azimuth domain, whereas Calterah’s solution leverages beam-range-Doppler features and a new radar signal processing flow to extract more frequency information, significantly reducing environmental interference and enhancing system robustness.

2. Lightweight CNN Models for Realizing Multiple Advanced In-Cabin Radar Functions

Calterah has developed two lightweight convolutional neural network (CNN) models, employing the resampling technique to address label imbalance, the label smoothing technique to reduce label noise, and regularization to enhance the generality of models. Using a lightweight CNN model, Calterah’s in-cabin radar solution achieves three advanced functions simultaneously, namely detection, localization and classification, with the improved radar system stability thanks to the enhanced generality, to cover more diverse detection scenarios.

3. Post-Training Quantization for Efficient Deployment and Operation

Calterah’s in-cabin radar solution utilizes the post-training quantization technology, reducing the model's memory footprint and speeding up inference. Specifically, in the solution, Calterah has deployed a deep learning model on the Rhine-Pro SoC for the first time, occupying only about 76 KiB of memory, with a single inference taking just 44 ms. This makes the solution easy to deploy while meeting the real-time requirements of in-cabin applications. The post-training quantization technology also enhances inference speed and reduces power consumption, providing more time and energy efficiency for other functions.

In his presentation, Dr. Zhifei Wang showcased a real-time in-cabin radar demo based on the Rhine-Pro SoC: the radar system stably and accurately detects the seat occupied by a baby model, even when covered by a blanket. It can also distinguish between adults and babies. When both an adult and a baby are in the vehicle, the radar identifies the adult and does not trigger an alarm, but when the adult leaves, the radar detects the baby and issues an alarm. Moreover, limb movements of occupants do not affect detection, localization, or recognition and vehicle shaking does not cause false alarms.

Calterah also showcased its latest imaging radar chip technology — the two-chip cascading radar solution based on Andes SoCs, at the Automotive Forum of EuMW 2024. During a panel discussion, Calterah joined international OEMs and Tier 1 companies to explore "Innovations and Technology Trends for Automotive Radar Devices Towards High Performance Imaging Radars."

At EuMW 2024, Calterah demonstrated how its cutting-edge mmWave radar chip technologies empowered in-cabin sensing and intelligent driving and also engaged in in-depth discussions with global industry experts. As the automotive industry embraces smart mobility, Calterah will continue leading innovations in mmWave radar semiconductors, helping to create safer, more convenient and more comfortable smart car cabins and intelligent driving systems