Principled Design Within Known Signal Models
Physics-based signal processing for model-driven sensor and algorithm development
Modern vehicles are equipped with numerous sensors. In the field of sensor signal processing, we focus on active sensors such as radar, sonar, and LiDAR, which measure the position, velocity, and physical properties of objects by utilizing the reflection and scattering of electromagnetic waves or sound waves. Signal processing spans a wide range—from low-level signal processing that estimates position and velocity directly from raw sensor signals, to higher-level processing that aggregates estimated information to recognize the states of concrete objects such as pedestrians and vehicles. Our research covers both. In recent years, machine learning has been applied at every stage of processing.
A key feature of active sensors is their ability to achieve a high signal-to-noise ratio by transmitting known signals. Beyond simply obtaining a high SNR, by carefully designing the transmitted signals, it becomes possible to directly acquire physical information that cannot be obtained with passive sensors such as cameras—for example, measuring distance based on time of flight (ToF) or directly measuring velocity using the Doppler effect. While the sensor hardware itself is developed at DENSO headquarters, our work is not limited to signal processing algorithms alone. From a signal processing perspective, we also consider the design of transmit–receive schemes required to achieve desired sensing performance.
Among active sensors, I work on developing environment perception algorithms using millimeter-wave radar. My work is relatively focused on later-stage processing, where point cloud data are processed using neural networks to detect objects. Before joining the company, I conducted research on event cameras (see our Computer Vision research section). Although both domains deal with time-series of point cloud data, I found the differences between them intriguing, which led me to become interested in this field.
Image recognition has become a mature, highly competitive field where large-scale data is increasingly the deciding factor.. In contrast, deep learning for millimeter-wave radar has not yet reached that stage. What I find particularly interesting is the opportunity to leverage radar-specific advantages—such as the ability to measure velocity directly—and to design methods that make the most of these unique characteristics.
In contrast, my work more often focuses on low-level processing rather than later-stage processing. For active sensors, it is essential that the signal characteristics are known in advance. This makes it possible to derive, from physical models, what kinds of observations can be obtained under various conditions and what level of estimation accuracy can be achieved. As a result, sensor design itself can be approached in a highly systematic and physics-grounded manner. From an engineering perspective, this represents a gold standard in engeneering—one of the ideal forms of engineering practice. I also find it fascinating that discussions can be grounded in physical principles.
The Appeal of Creating Countless Application Domains from Measurement Technologies
Expanding active sensing applications from vehicles to factories and beyond
Active sensors have long been used in automotive sensing. For example, sonar is used to measure distance and issue warnings when a vehicle is about to collide with an obstacle in a parking lot, while radar is used to measure longer distances and apply braking when a collision is imminent. Vehicles are equipped with numerous sensors, and by making full use of them, technologies for safe driving assistance are brought to life.
There are also approaches that attempt to achieve autonomous driving using cameras alone. While this may be possible, active sensors remain essential when aiming for greater safety and robustness. Cameras require sufficient light to capture images, whereas active sensors emit their own signals and can acquire information regardless of brightness. Imaging radar, which increases resolution by arranging multiple antennas, can even provide high-resolution images similar to those of cameras. Because vehicles often operate at night, complementing cameras—which are vulnerable to low-light conditions—with active sensors is a key enabling technology for autonomous driving.
In general, sensors are technologies for “measurement.” Measurement is indispensable in engineering, and it is essential to evaluate whether our intended outcomes have in fact been realized, and if not, to identify what went wrong and improve it. In this sense, the application scope of signal processing is virtually limitless. For example, to implement a function that warns when a baby is left inside a vehicle, we must first carefully observe and understand what is happening when a baby is present in the cabin. There should be multiple signals—such as breathing or temperature—and we consider whether it is possible to extract them from noise and disturbances, and then develop algorithms accordingly.
Beyond vehicle functions, there are also many “measurement” challenges in factory environments. In automotive products, new functions are typically released in cycles of four to five years, which means it can take time before research results are implemented. In contrast, in factories, research outcomes can sometimes be deployed on much shorter timescales.
Researching Both Real-World Problem Solving and Highly Generalizable Core Technologies
Bridging product-driven challenges and foundational signal processing research
One of the strengths of ITLAB is the ability to conduct research close to real products. As an automotive parts manufacturer, there is the excitement that one’s research outcomes may eventually be installed in vehicles, and it is also possible to conduct research that addresses real-world challenges unique to manufacturers. At the same time, we can also study highly generalizable core technologies with broad application potential and publish academic papers. Researchers here have the freedom to pursue both.
My current research on millimeter-wave radar is very close to product implementation. I hope to overcome practical constraints such as computational cost and, in collaboration with DENSO engineers, carry the work through to actual deployment in products.
What I would like to work on going forward is building models that enable sensor performance evaluation at higher levels. For low-level processing such as distance estimation, sensor performance can be evaluated using physical models. For example, in radar systems, once requirements such as distance resolution and velocity estimation accuracy are specified, physical models determine the required hardware specifications, including the number of antennas, sensitivity, and transmitted power. However, even if performance requirements for pattern recognition in autonomous driving are defined, the task complexity makes it challenging to theoretically determine the necessary hardware specifications. I would like to develop principles that bridge this gap—connecting the inherently uncertain world of neural-network-based pattern recognition with the classical world of signal processing.
One of ITLAB’s major strengths is that, as part of DENSO, a company deeply engaged in product development, there is an abundance of product-driven challenges and opportunities to apply a wide range of technologies—meaning there are countless research themes. We also have access to data that are not part of publicly available datasets. For those who are interested in tackling problems that are not yet visible to the outside world, this is a highly rewarding environment.
Join Our Research at DENSO ITLAB
Where engineering curiosity
meets real-world impact
