COGNITIVE AUTOMOTIVE RADAR
Radar systems are becoming increasingly smaller and less expensive and their software-controlled sensors allow completely new sensing strategies and signal processing algorithms that are adaptive and capable of learning from experience. Hence, they can revolutionize modern driver assistance systems and, with new sensors, can also pave the way for autonomous driving.
Thanks to the development of semiconductor oscillators and amplifiers, more efficient switching between transmission and receiving modes as well as digital waveform generators, circuit technology for millimeter wave radars has made great progress in recent times. Basic functions such as mixing, filtering and modulating can be controlled by software modules and are no longer strongly dependant on hardware. This clears the path for compact, low-cost multi-channel radar systems that are extremely versatile and highly reconfigurable. These advantages favor the design of new signal generation and processing algorithms which greatly facilitate on-the-fly optimization of radar operations. Also very promising are the software-controlled radar systems, above all for automotive applications, where the adaptability of the radar in terms of bandwidth, measuring times or channel allocation is decisive for reliable information acquisition from the respective environment.
Automotive radar systems have to adapt to various scenarios. City traffic is, for example, characterized by high target density and diversity which have to be reliably detected despite the highly heterogeneous background. Challenges on the highways, on the other hand, include the detection of targets traveling at high speeds and long-distance detection. Driver assistance systems should be flexible and equipped with functions such as active speed control, lane-keeping assistance or pedestrian detection. This necessitates the high-precision estimation of short and long distances, relative speed as well as the angular position and resolution of several different targets in the relevant environment.
The work group "Adaptive Perception", which was set up to address these challenges, conducts research for its customers and partners from the automotive industry on array design for multi-channel systems and innovative digital beamforming techniques. The group aims to improve spatial resolution to such an extent that targets moving at the same distance and at the same speed can be reliably detected and distinguished. Research is also conducted on radar systems that can adaptively adjust their sensing and processing operations so that their performance can be improved over time.
Cognitive radar scans its environment in an adaptive manner. It uses information from previous measurements, such as the number of targets, their parameters, the noise level and clutter distribution as well as knowledge from other sources, e.g. other sensors or cartographic databases, to adapt its own parameters. A cognitive system normally comprises four elements (Fig. 1): the illuminated scene, the sensor, a processor, which processes the raw data and estimates the scene parameters thus generating a perception of the scene, and a cognitive controller. This dynamically optimizes the system's operative parameters and allocates the resources for the next measurement.
One challenge associated with spatial resolution is the transmission and reception of the channel selection for position determination. In contrast to distance and speed determination, where the bandwidth, measurement time and measurement count are decisive to resolve nearby targets, the accuracy of angle estimation depends on the length of the antenna array. Accuracy increases proportionate to the length, i.e. the number of antenna elements in the array. These elements must, however, be adequately distributed to prevent aliasing and would require additional Tx and Rx modules with large volumes of data for processing which would ultimately result in significantly higher system costs. For this reason, the scientists at Fraunhofer FHR developed new compressed sensing algorithms which can also extract information from sparsely equipped antenna arrays and have already demonstrated super-resolution characteristics.
Moreover, the current focus lies on the adaptive selection of the receive channels and transmitter activation sequences through time-division multiplexing to facilitate cost-effective multiple-input multiple output (MIMO) operation. A Bayesian filter is used to sequentially include previous measurements in the probable distribution of the angle position. Through the utilization of complex metrics to optimize the channel selection for the next measurement, the scientists quickly arrive at a value that is close to the actual value of the angle position (Fig. 2).
In the next step, initial experiments will be carried out with cognitive MIMO and the newly developed algorithms will be enhanced with the ability to develop and use previously defined measurement strategies. This is particularly useful in scenarios that require foresight or forecasting. The solutions created by the work group help companies to take full advantage of the unique sensor properties for efficient and safe driving.