Hello everyone, I am electric small detection! Xiaopeng is about to X-Martos2.6 version, give Zhun version of the owner OTA a new function – parking memory parking (referred to as VPA), it is said that the owner will not have to open itself, vehicle You can find a fixed parking space completely. I have a “electric vehicle is the principle of a bad disaster”. How can I easily let this functional opportunity? Not much to say, let’s talk about Xiaopeng’s VPA today.
VPA Hardware Architecture AMP; Technology Directions
Before I interpret the hardware, I will distinguish the name and face it. The industry is generally called “AVP Matchbolo”, the alphabetical order is different, but it is not a mistake.
The article written in a small detective, 3 of the industry AVP technology routes have three. One is to transform the parking lot, install the sensor to the parking lot, tell the vehicle “how to go”, call the field route; the second is not to transform the parking lot, through the vehicle own camera / radar / positioning technology to achieve self-decision-making and Control, called the car route; Article 3 is that the host factory cooperates with the automatic driving company to help the vehicle make behavior decisions through local and cloud real-time communication, this calling the cloud cloud.
Xiaopeng VPA uses the car technology route, and the full stack of autism, does not rely on suppliers, and does not rely on parking lot, but develops to the main plant. The capacity requires very high. Dare to play this, mainly P7 itself has a very rich sensor hardware, and the Xavier platform can meet the needs of VPA to environment.
However, underground parking is a relatively extreme environment, light, social vehicles, obstacles, pedestrians, are uncertain, far more complicated than closed roads. . We usually say that the sensor category includes a camera, millimeter wave radar, ultrasonic radar, where millimeter wave radar is complicated in the ground due to the complex route, which is basically unavailable from it.
Therefore, Xiaopeng P7 calls all 14 external cameras in the whole car, involved in environmental perception and construction. In addition, 1 of the driver’s attention is detected with a support of the steering wheel infrared. The external camera also has a different division of division: 
Before the main photo and telephoto camera, the external rear view, participate in the identification of 50 meters away from the object;
Two fish eye cameras Responsible in the middle distance, cover 360de; range 3 ~ 100 meters;
4 360DEG; panoramic view camera is mainly used for parking, detects a close-range environment information of 0 to 10 meters.
Of course, 12 ultrasonic radar acts as a visual aid, which can detect close objects. In addition to these sensors that our naked eye, Xiaopeng P7 has a cost-not-low thing -IMU inertial measuring unit. The IMU can calculate the position of the vehicle in real time by detecting the angular rate and acceleration of the object, and high accuracy is high. The gap has no GPS signal, so the real-time positioning of the vehicle is basically completed by IMU.
Many P7 owners have this experience: driving in the library, even if the GPS signal is gone, the large screen map still shows the real-time position of the vehicle very accurately. This is achieved by IMU. I have driven from get off work, and my wife can use the app to see the accurate real-time position of the vehicle, even when the side of the head can display it, properly check the tool.
How is the actual experience?
Xiaopeng’s parking lot memory is currently only open to open offline garages, like NGP, the owner is after reading the video and can be enabled after testing. VPA supports vehicle learning of 100 paths, but it is implemented in daily life, the highest frequency of use is the company and the community. 
Before using the VPA, the owner needs to open a parking route from yourself, and manually complete the entire parking process. VPA’s interaction is high, and the vehicle can demonstrate the learning environment and route in real time on the big screen, and the display effect is quite smooth. Learn againAfter the owner, the owner took the car out of the base, and the vehicle will re-access the GPS signal as a basis. When returning to the starting point again, the VPA can officially take over the vehicle, complete the duplication and parking process.
VPA in subsequent use, vehicle integrated IMU and ground markers, such as columns,, to overlap these real-time data with learning virtual space In the case of losing the GPS signal, the centimeter position is achieved. Here Xiaopeng himself did a test, placed two row of paper cups on the path, and the single side width only exceeded 10 cm of the wheel, and the path error of the VPA driving was small to the paper cup.
Here there is a few questions to explain with you. 
How does the vehicle judges that the land is B1 and B2 or B3?
Through the IMU, it is possible to calculate how much meter of the vehicle vertical distance is reduced, thereby judge the number of layers.
Why don’t you support cross-layer looking parking spaces? Why is the open air parking lot? 
Sprinkler parking, the ground parking lot will be open, don’t worry, Xiaopeng will complete the full scene VPA by step by step, after all, have to consider the owner’s learning and acceptance process.
In the car experience VPA, the vehicle speed is slower than yourself, basically no more than 18km / h. Due to the study process, the vehicle has also been recorded by visual and iMu, so it is again facing the speed tape, and the vehicle will reduce the vehicle speed, avoid serious bumps in the car. At the same time, interaction is still a very colorful place in Xiaopeng. In the VPA state, the voice assistant small P will frequently prompt the driver’s vehicle, such as steering, pedestrians, social cars. In the face of the intersection, flashlight, the steering lamp is facing these gauges.
Human meat test begins!
Because the entire VPA experience process is too smooth, wave is not shocked, I decided to switch the role, get a pedestrian from the driver to see if the VPA is not enough in the eyes of the pedestrian. Below this video, strongly recommended that everyone will watch, 1 minute high-energy warning, small detective, rope, dangerous, carwhat!
(Details Operation Video Click here)
In order to simulate the trailing route without regular travel route, I tried to stand still blocking the route, respectively. Several different bursts such as crossing the fence crossing, ghost probes. I used my own intact legs to indicate that P7 has done a successful identification of pedestrians and brakes. The most difficult probes are the most difficult probes, but P7 is still successfully stopped when it is less than half rice.
Everyone can also see from the above animation, how fast my ghost probe came out. The system quickly responds, and the owner takes the vehicle through voice and large screen. I feel that I can’t “die” again, if I changed into a human driver, I will definitely get off the car. I am …
Already finished, VPA height Relying on the camera, I decided to use the mud to stain the camera, see if P7 can complete the memory park if the P7 is blocked in the sight. For the first time, I stain the glass of the three-purpose camera, effect similar to the fog of glasses. After the results of the result, the VPA can still work normally, and the behavior of the vehicle seems to have no difference, and it still has a batch.
On the first time, I put the 360deg of the bumper; the front view camera was pasted with mud, blocked nearly half of the line of sight. As a result, the vehicle still completed the memory park with a 360deg camera and a 360deg; the camera was disturbed. Unfortunately, I have no courage to simulate pedestrians in this case. Personally think that Xiaope’s VPA is relatively balanced in sensor fusion, algorithm, and positioning. It will not depend on a sensor. Otherwise, the camera is inseparable, and it will greatly affect the owner’s experience.
Current VPA can accurately identify important elements such as vehicles, columns, deceleration belts, and pedestrians, which will also support identification brakes, pets for the parking lot. , Tricycle, isolated pile, etc. In addition, the VPA’s learning is not only once, in the use of the same path, the vehicle will continue to optimize the route and decisionSection. In the future, I believe that VPA will also use the functions that use it.
Although the valet parking will not become a key factor in the decision of the car, it can be seen from the back of the technology. Xiaopeng NGP is expanding from a single high-speed closed environment to a complex low-speed open environment. Thanks to the full stack of self-research route, Xiaopeng can thoroughly dig the sensor and chip power based on the existing sensor, which is the foundation of the VPA has a stable experience.
Back to the beginning of the parking technology route, this year we will see the car solution represented by Xiaopeng, and the car field representative of the Weima AMP; Baidu is a car. I believe that the more significant meaning of valet parking is that automatic driving interactions from the vehicle-driver, expanding the vehicle – driver – foreign personnel. The big mom danced by the local Turi Square Dance knows that the car can be found in the car, and the automatic driving is truly intended to gather in the field of vision. From now on, the car blows the angle of the capture of “last 1 km”.