In August 1888, Mercedes-Benz founder Karl Mercedes-Benz's wife Bertha Benz opened her husband invented three rounds of internal combustion trucks from Germany Mannheim to Pforzheim, this time approximately 100 kilometers of history marked the birth of modern cars.
125 years later, in August 2013, the Mercedes-Benz took over the past, but this time is no one driving, in order to pay tribute to the sages of the seniors, Mercedes-Benz will have a basic non-modified Mercedes-Benz S500 Kinddom called Bertha.
The core sensor of Bertha is a 1024 * 440 pixel FOV of 45 degrees, and the Mercedes-Benz is increasing to 35 cm long, mainly in order to cover further distance, and the standard version S-class bib is 30 cm. Both SGM and rod pixel calculations are completed using FPGA, and 400,000 independent depth measurements per frame can be used. The remaining calculation Mercedes-Benz recognizes the traffic light with a single-grade of 90 degrees, and uses an FOV 90 degree single-grade to make a feature-based auxiliary positioning. There are also 4 120 degrees of medium-distance millimeter-wave radar, and KIT provides high-precision maps, including speed limit, zebra-line position, stop line, road curvature, and the like. Because it is a suburban, Mercedes has only used GPS, and does not use an external inertial navigation system. This is also the closest to mass production L4 unmanned car.
The double object has an overwhelming absolute advantage, which can be done, and the binocity can do, but the single-minded 3D stereo vision. Taking traffic detection as an example, the video of Mobileye red lights must have seen it. Identify the red green light is one of the most difficult technical points in the perception. Baidu and Google are using the respective street view stock resources, using a priori information. Customize ROI to improve the accuracy of identifying traffic lights. However, the update of street views is slow, and China is inappropriate in rapid development, it is suitable in the United States. Even so, Waymo also appears over the red light video. If there is no street view, simple single-purpose traffic light recognition rate is quite low. More deadly, my country's traffic lights are tit, especially in Tianjin, there are five traffic lights on one road, just a nightmare of unhocusebrus. And V2X is still far from time.
There are currently three two-dimensional identification traffic lighting method, the first is to use the target candidate zone filtering method, analyze the visual difference in the target candidate area, separation of the foreground and background, no prior knowledge.
The second is associated position filtering method, this method requires a priori knowledge, which is the 3D position information of the statistics traffic light, especially the height of the traffic light, and the straight distance from the zebra line, linear distance from the road edge. The three-dimensional distance information of the target can be obtained when the drone is on the line, and the priori kingdom will be filtered. This method can greatly increase the red green light recognition. Sometimes only the high information of traffic lights, relatively street views, this prior knowledge can be easily obtained.
The third method, that is, the projection method. Model the real world of traffic lights, and increase depth measurement data, the red green light is to prohibit the red green light model, and the re-cooperation of depth measurement data is not a traffic light. No matter that, it is much better than a single look.
The three-dimensional matching of the double destination is almost the completion of FPGA. It is very rare to understand the algorithm and understanding of FPGA. The price of FPGA is also relatively high, which limits the binding application. Plus FPGA manufacturers are located in the United States, and Chinese companies are concerned about this. The United States sells one eye to the FPGA, once it is true, it is likely that the computing unit exceeds 100,000 FPGAs. Single-grained applications are hardware, the price is low, so the application is far more wide.
However, this phenomenon will have a large change in the next few years. The two major three major automotive processor manufacturers in the world launched a dedicated binocular design processor, which is the S32V3 series of Rissa's R-Car V3H and NXP. Both are samples in the third quarter of 2018. Both can not have FPGA.
Let's take a look at Rya's R-Carv3h.
R-Car V3H has been determined to obtain Nissan and Toyota orders, Nissan will use R-Car V3H in the whole line, including highway automatic driving, remote parking, automatic parking, plug-in assistant, unrestricted AEB (L2 stage) The limit conditions are limited.
R-Car V3H operates to 4.2TFLOPS, more than 3TFLOPS of Mobileye Eyeq4, manufacturing, R-Car V3H, with absolute advantage, TSM 16 nm FinFET process, semiconductor 28 nm FD-SOI process. Of course, Mobileye Eyeq4 is two years compared to R-Car V3H, but R-Car V3H has a powerful CPU system, including 4 A53 and a R7 with lock-up function, which means that R-Car V3H should meet ISO 26262 ASIL-A or B standard rather than Mobileye can only pass the AECQ-100 1 standard.
The above picture shows the internal framework of R-Car V3H
R-Car V3H For the L3 +-grade vehicle, R-Car V3H contains three-dimensional parallax and optical hard core IP, which is equivalent to FPGA, perhaps slightly better than FPGA.
NXP plans to launch S32V3 samples in the third quarter of 2018, and V3H is also a stereo parallax and optical hard core IP.
The S32V3 is up to ASIL-D level, and the security level is far more than possible.
S32V3 internal frame map
In addition to NXP and Rissa's top companies, there are also products of driving recorder chip giants, computing performance than Mobileye will be strong in the third quarter of 2018. In 2015, Anta acquired Italy's startup Vislab in 2015, which was founded by the University of Parma, Italy. At present, the founder of Alberto Broggi serves as the person in charge of the Antiba automatic driving business.
Anaba first generation automatic driving chip CV1 can correspond to two 8 megapixels of two-purpose stereoscopic vision, and the operation has reached 2TFLOPS. With Sony's IMX317 image sensor, the pixel accuracy can reach 3840 * 1728, such a high pixel, even if the FOV is 75 degrees 30 cm baseline, the effective distance of 300 meters can also be achieved, and the effective distance of the large-over larger laser radar.
The effective distance of laser radar is extremely close to the object reflectivity. Usually, the manufacturer only gives an effective distance when the reflectance is 80%. For white vehicles, the reflectance may only be 10%, and the effective distance will shorten to 80%. 1/3 Even less. Generally, the MEMS laser radar is only 30 to 70 meters at 10% reflectance, and the mechanical rotation type is slightly better.
CV1 can directly output a parallax map, and the frame rate is one frame per second. At the same time, it also contains roadblocks or isolation gate detection, road edge and lane detection, traffic signal detection, general obstacle detection.
CV1 is only an anti-test water, and the Anta launched CV2AQ in early 2018, the operation performance is 10 times, approximately 14TFLOPS, 10 nm technology manufacturing, through the AEC-Q100 2 standard.
This chip is abnormal, can handle 32 million pixel data at the same time, can be correspondingly 6 bidders, including two 8 million pixels, and 4 2 million pixels. The CV1 can only correspond to one 8 million pixels.
With the bottleneck of the FPGA, the large-scale application of the double is coming! , Read full text, original title: Break through the FPGA restriction, the two-purpose perspective vision era is coming soon
Article Source: [Micro Signal: Zuosiqiche, WeChat public number: Zissi Automotive Research] Welcome to add attention! Please indicate the source of the article.
Our other product:
