TU-TESL

• 22

Automatic parking forward parking space detection, more accurate LCC (Lane center assist), ALC (automatic lane change assist) and traffic light recognition The new functions on the p7 of Xiaopeng are emerging one by one. As Xiaopeng's second mass production vehicle, xpliot 3.0 on P7 adopts a new hardware architecture, especially in the sensing layer. What does comprehensive self-study mean? Among the independent R & D vehicle enterprises, only one has been able to establish the perception ability completely. It is not a small challenge to break the monopoly of mobile ye, a giant of ADAS, and become the second car company with self-developed perception ability. Therefore, how to deliver more intelligent and experience better automatic driving functions to users is also the core thinking of Xiaopeng automobile's research and development of automatic driving. To build a full closed-loop self-research capability "to provide end-to-end full closed-loop self-research and development capabilities to adapt to China's road conditions. This was put forward by Xiaopeng in October 2019. What are the so-called closed-loop self-study capabilities? Xiaopeng automobile is also exploring the puzzle of the answer: starting from the perception, decision-making and implementation, the hardware architecture, application layer and interaction layer of automatic driving are comprehensively redesigned. Xiaopeng automobile has invested a lot in its own research, from the central control panel, bus, electronic and electrical architecture, to the hardware architecture, and so on. All the internal teams have participated in the design. This year, one of the most obvious changes is that the supplier's scheme is adopted in the development of G3's automatic driving perception scheme. However, with P7, Xiaopeng began to carry out thorough self-study and reconstruction of the vehicle's perception system, so as to achieve more accurate target perception and better data collection effect. On this basis, Xiaopeng team has developed a number of localized automatic driving assistance functions suitable for domestic road conditions. "Completely self-developed" and "Chinese characteristics" have also become the key words of Xiaopeng's automatic driving technology. The advantages of self-study are self-evident: first, it can develop the function of travel alienation; second, it can speed up the function iteration. In other words, if Xiaopeng automobile, which regards intelligent driving as its core competitiveness, wants to promote the mass production of automatic driving, it must solve the localization adaptability problem and push the user experience to the extreme. The forward parking space recognition function is the product of this thinking. On the xpilot 3.0 system of P7, the forward parking space recognition of Xiaopeng Auto Park is a unique function in the industry at present. Forward parking space recognition is to use the side forward camera to find the front parking space in advance. The CEO of Xiaopeng automobile has said that Xiaopeng hopes to increase the use frequency of auto parking to 70% - 80% or even higher. Therefore, the team has invested a lot of R & D resources to optimize the experience of auto parking, which makes Xiaopeng's experience of auto parking better than other models of the same level. In this context, it is necessary to build a complete independent R & D system. Known as the strongest "xpilot 3.0 hardware architecture" in the industry, compared with the hardware configuration of Xiaopeng's first G3 model, users can easily feel the team's intention on P7 equipped with xpilot 3.0. Xiaopeng P7 has added these key components in hardware: what is the concept of integrating these configurations? On the computing platform, Xiaopeng chose NVIDIA's Xavier chip, which is the first car gauge chip used in mass production vehicles in China. How powerful is Xavier? The second generation hardware of Tesla autopilot (hw2.0 / 2.5) is equipped with NVIDIA drive PX2 chip, and the next iteration product of PX2 is Xavier. Xavier's computing power reaches 30 tops, which is the most powerful commercial vehicle computing platform in the industry. Even for the eyeq5 (which is expected to reach 24tops) to be mass produced by Mobileye this year, P7 still has a small advantage in computing power. In addition, xpilot 3.0 also launched the fifth generation millimeter wave radar of Bosch, the world's top auto supplier. Compared with the fourth generation radar, the perception performance of the fifth generation radar is improved in an all-round way: the horizontal field angle is increased from 90 ° to 120 °; the point cloud density is increased by four times, which is more than that of the current Tesla Model 3 radar; at the same time, the detection distance, the number of objects and the accuracy are also increased correspondingly. The configuration of 5 MMW radars and 13 cameras also enables P7 equipped with xpilot 3.0 system to complete 360 ° dead angle free coverage of the whole vehicle, thus ensuring the intelligent function and omni-directional perception requirements of P7. With the complete self-study of the sensing layer and the latest sensors in the industry, according to the insiders of Xiaopeng automobile, xpilot 3.0 architecture is the industry leader in 2020, and there is no problem that it will continue to be the leader in two or three years, and this architecture also provides sufficient support and space for the subsequent evolution. In addition, xpilot 3.0, which adopts a new hardware architecture, takes the computing unit as the core. The biggest difference between the two hardware architectures is whether they have a central scheduling system. According to Huang Xin, Xiaopeng's auto driving product director, the benefit of this approach is that it makes the central brain have strong scalability and compatibility. For example, some functions of automatic driving may not be used in the first and second year, but OTA push can be carried out to activate these functions in the third and fourth year. Take the NGP (navigation guided pilot) of xpilot 3.0 of Xiaopeng as an example, which is the automatic driving function set of Xiaopeng automobile. NGP is similar to Tesla's NOA (navigate on autopilot), which includes the functions of autopilot in different scenes. In theory, NGP not only includes LCC, ALC, TJP and other conventional automatic driving assistance functions, but also includes the differentiated functions built by Xiaopeng automobile for localized roads. The structure of traffic light recognition also determines that NGP will be an evolving product, and P7 with xpilot 3.0 will continue to get more automatic driving functions. In fact, in order to realize the L2 + level automatic driving with better experience, the architecture similar to P7 is basically adopted. The reason why Tesla can keep Ota is that it benefits from the flexibility of architecture and decouples a large number of systems, so that each system becomes an independent controllable unit. The function iteration of Xiaopeng P7 has a new hardware architecture, and independent research and development can quickly launch differentiated innovative functions. Then, what are the improvements of the automatic driving assistance function of P7 compared with G3? In late March this year, the R & D team of Xiaopeng automobile once drove the p7 engineering vehicle to carry out a test on guanfo expressway. From the video of lane line recognition, it can be seen that when passing a section of road near Nansha toll station, complex road conditions such as superposition of residual lane line and normal lane line, vestige of visual deceleration line and so on appear. However, P7 passed the section smoothly under the condition of LCC function, and was not interfered by the wrong lane line. The conventional method of lane line detection is to identify and judge the lane line by the perception system. This method will encounter problems that can not be dealt with when encountering complex situations such as lane line superposition. P7's strategy is to train the data in the special case of lane line superposition, and then use the method of deep learning to effectively eliminate the interference. In addition, in G3, when the ALC function was originally turned on, considering safety factors, G3's lane changing speed was very slow. On P7, because of the improvement of hardware perception, users will feel more comfortable when using lane change. According to official information, more accurate functions such as LCC, ALC and automatic parking forward detection will be opened on P7 in the first half of this year. In the second half of this year, xpilot 3.0 system of Xiaopeng P7 will open 360 degree ring body perception system, and use 360 degree perception ability to release more interesting functions. Now, P7 reshapes the hardware architecture, which is equivalent to consolidating the cornerstone of the flexible update of automatic driving. Xiaopeng automobile chooses the self-developed route, which provides guarantee for its model to build the localization function adapted to the domestic road conditions. This is also the competitive power of the new car making force which is good at software development, which is different from the traditional car enterprises.