From September 14th to 15th, 2022, Wuhan Nordson Co., Ltd. will hold an automotive radar forum in Anting Automobile City, Jiading District, Shanghai. The forum mainly focuses on the development of radar chips, the development trend of radar technology, test verification, mass production, system design, and components As well as material development, etc., the forum speech topics are as follows.
Participate in the forum on September 15th to follow up the progress in the field of radar at home and abroad. Listening to the reports of many experts throughout the day, I mainly have the following feelings.
1. Radar systems are getting more complex
In the past few years, the radars were generally 2T4R and 3T4R. Now leading foreign radar manufacturers have mass-produced 12T16R, such as the Continental ARS540. Arbe and Mobileye are currently developing 48T48R, with 2304 virtual channels and angular resolution close to laser level.
The increase of array antennas first brings about the complexity of RF design, and secondly, the data volume of signal processing also increases sharply, requiring higher-performance processing chips (partially using FPGA), and the space on the car is limited, and the structural design, heat dissipation, etc. higher requirement.
In addition, the increase in the number of channels also brings greater challenges to the production process and production line testing.
Overall, there is a big gap between domestic manufacturers in this regard.
2. Subversion of radar algorithm
Radar algorithms are basically divided into signal processing algorithms and data processing algorithms. Signal processing is from ADC data collection to output point cloud, and data processing is from receiving point cloud to output target.
The signal processing algorithm can use the classic multi-dimensional FFT+CFAR, but the current trend is to put the signal processing algorithm in the chip, so that the operation efficiency is higher. The application only needs to configure parameters, which greatly simplifies the algorithm development process. But engineers who do signal processing algorithms may be out of work.
In addition, most of the data processing algorithms used the 2D point target tracking algorithm before, but the current 4D radar outputs 3D point clouds, and the data processing algorithms need to consider new frameworks and methods. The 3D tracking model of lidar is a feasible direction. At present, there are many papers using radar point cloud for target detection, recognition and tracking, mainly abroad, and relatively few in China. Radar manufacturers have insufficient accumulation in this area, and there is still a long way to go.
3. Radar performance gap is obvious
The radar point cloud of Mobileye's 48T48R is very stable, and the follow-up target detection and tracking effects are also good. However, the domestic manufacturer 6T8R’s radar point cloud has a close range and large fluctuations. It may be that the front-end RF phase noise is high, which makes the detection effect poor.
The picture below shows Mobileye's point cloud and detection results, which match the actual object very well. Such hardware and excellent algorithms are naturally expected to achieve perfect results.
In addition, in addition to the conventional method of increasing the angular resolution by increasing the number of arrays and channels, Zerdar proposes to use the method of synthetic aperture, by installing multiple 3T4R radars at different positions on the vehicle, and directly using the results of RV-map with machine learning To detect the target, to improve the ability to distinguish the target.