Shulou(Shulou.com)11/24 Report--

As a drone company, DJI achieved its head in its own track and became the strongest drone company in the world. This resume also makes DJI one of the domestic star technology companies, and every major change will attract the attention of the industry.

In April last year, DJI officially launched the DJI car brand and began to make every effort to lay out intelligent driving. As the cross-border layout of giant enterprises, DJI vehicle has become the object of attention of the industry at the beginning of its release.

On the other hand, with DJI's deep accumulation in software and hardware, many people in the industry are very optimistic about DJI's cross-border.

A year after the launch of the brand, Dajiang finally handed in its papers, and the KiWi EV model developed in cooperation with SAIC GM Wuling demonstrated the auxiliary driving ability in the production car for the first time.

▲ KiWi EV

And Dajiang's way of playing is also different from many domestic self-driving companies. Dajiang does not take the strategy of multi-sensor, high computing power and high cost, but chooses a low-key and pragmatic road, and the cost of its L2-level auxiliary driving is very low, so it landed on a mini car-SAIC GM Wuling KiWi EV for the first time.

More importantly, although it is mainly based on visual perception, Dajiang does not choose monocular vision like other enterprises, but uses binocular vision, which avoids some disadvantages of monocular vision, such as inability to identify still objects, and makes up for the large amount of computation of binocular vision.

From this point of view, DJI has accumulated a lot of foundation in the field of intelligent driving, and it also shows its ambition to seize the market quickly.

So, what is the performance of this system? What difficulties have been overcome in its development process? What's the next step? After testing and communicating with the car engineers in DJI, the car found the answers to these questions.

01. The performance of the bend with large curvature is amazing in response to the immaturity of Gasser. Let's first talk about the details of the field test.

At present, DJI's intelligent driving products are mainly D80, D80 +, D130, D130, of which D80 and D130 are mainly based on L2, including ACC automatic cruising, automatic car following, automatic parking and other functions, while 80 and 130 represent the maximum speed available to the system. D80 + and D130 + refer to the advanced version of assisted driving, with the addition of lidar, driving behavior identification and early warning camera, and four look-around cameras, which can achieve quasi-L3 intelligent driving.

The vehicle tested this time uses the D80 system, that is to say, the KiWi EV model can only start auxiliary driving at a speed below 80km / h. So the tests are mainly focused on urban areas and urban expressways.

After learning this information, the car did several targeted tests.

When ▲ presses the button, he can turn on the intelligent assistant driving with one button.

1. Drive in the middle of the lane

The first is the single lane center function, for the basic L2, this function is very basic, but it is very difficult. Many car companies will "draw snakes" after opening the single lane center.

On the other hand, the system of Dajiang gives a very good impression on the things of the car, and the vehicle can be kept steadily in the middle of the driveway. In the absence of a lane, the vehicle can also follow the car in front of the intersection, after which it will continue to keep the lane in the center.

▲ DJI L2 lane is centered and maintains a very stable performance.

2. Through the bend with large curvature

It is worth noting that the vehicle can not only maintain the central form on the straight road, but also pass very smoothly in the curve with large curvature, and the vehicle can also maintain a speed of more than 60km / h.

▲ Dajiang L2 successfully passed the bend with large curvature.

3. Follow the car at a low speed

In addition, the adaptive cruising ability of the vehicle, DJI this system gives me a very good feeling, especially in the traffic jam section, the vehicle can follow according to the driving situation of the car in front. After the front car stops, the vehicle will stop voluntarily, and after the front car starts, the vehicle will also start. You can even follow the car directly through the traffic lights.

▲ Dajiang L2 followed the car through the traffic light intersection.

This function can greatly liberate drivers and improve driving comfort in the scene of traffic jams.

4. Deal with the plug

In the process of testing, the car also experienced the ability of the vehicle to cope with congestion. In most situations, if the vehicle in front is diverted or jammed, the vehicle will slow down according to the safe distance ahead, and the whole deceleration process is relatively natural, and there will be no sudden stop.

The official video of ▲ shows that the water on the car of DJL 2 will not spill when dealing with the plug.

However, there are also some extreme cases in which vehicles do not respond well enough. If the vehicle is in the normal driving process, when the side car comes quickly from the back of the side and is jammed, the response of the vehicle appears to be immature and there is no obvious deceleration. The car found such a scene twice in the course of testing, and in the end, it was a risk aversion in the case of active takeover.

So on the whole, this set of L2 in Xinjiang still has an untimely response in the face of a more urgent situation, but as an L2 autopilot, human drivers have the responsibility to take over actively to avoid accidents.

▲ DJI L2 has insufficient coping ability in some scenarios.

5. Dial the pole to change the lane

Interestingly, the car adds an advanced function of changing lanes with a plucking lever. As long as the driver presses the turn signal, the vehicle will automatically detect the road condition in front and rear, and then decide whether to turn according to the road condition. When the adjacent vehicles are far away, the vehicles will accelerate to change lanes, while when the side cars are closer, the vehicles will give up changing lanes and wait for the right time to change lanes again.

▲ Dajiang L2 has the function of dialing the pole to change lanes.

At the ramp entrance, after pressing the turn signal lever, the vehicle will also automatically enter the ramp and pass through the ramp.

Although there are no stationary vehicles, special-shaped vehicles and accident sections in the testing process, according to the binocular vision algorithm adopted by KiWi EV, this technology can identify stationary objects more accurately and should also be able to identify stationary vehicles.

After an hour and a half of experience, the car feels that this system in DJI can almost meet the needs of assisted driving in the city, and the current performance of several functions is also good. According to the on-site engineer, a large OTA will be completed after the delivery of the vehicle, and the auxiliary driving ability of the vehicle will be further improved.

02. The cost of solving binocular algorithm problems is lower as the first mass-produced product after DJI entered the field of intelligent driving, many designs of this autopilot system are very interesting.

First of all, the product price is relatively low, sensor, using a forward-looking binocular camera + 4 look fisheye + 1 forward millimeter wave radar + 12 ultrasonic radar, and did not choose a large number of stack sensors.

▲ DJI L2 sensor configuration

Among them, forward-looking binocular cameras and fisheye lenses are produced by Dajiang itself, while millimeter wave radar and ultrasonic radar are purchased from parts suppliers. Among them, the most expensive cameras are produced by DJI itself, and DJI is very good at cost control, so the cost of the whole set of sensor configuration is not high.

In terms of chips, Dajiang is also the choice of outsourcing, but did not disclose its specific suppliers and product models, only know that its computing power is 16TOPS.

Similarly, DJI did not disclose the final price of this set of products, saying only that compared with other options, the cost of this scheme is very low.

Secondly, DJI uses binocular vision algorithm, but at present, there are not many domestic companies using binocular vision algorithm, most of which use monocular vision algorithm.

Monocular cameras need a lot of data feeding, and are constantly updated and maintained. all kinds of long-tail scenes need to be learned and tagged, but in specific work, monocular cameras need to first identify the types of obstacles and then avoid them. This characteristic has a time lag, which also leads to a number of traffic accidents.

The binocular camera is relatively close to the human eye, mainly through the parallax calculation of the two images to determine the distance. In other words, the binocular camera does not need to know what the object is, it only needs to identify the object in front of it, which can not only improve the accuracy of recognition, but also measure the distance accurately through calculation.

▲ DJI improves binocular vision algorithm

However, at present, not many companies use binocular cameras, which is mainly due to their high hardware stability requirements, large amount of calculation, close detection distance and other problems.

And how does DJI deal with these problems?

Binocular cameras are more complex than monocular cameras, so they are more demanding in module design and need to be certified at the vehicle specification level. DJI has a very deep accumulation in hardware design, has shipped tens of thousands of intelligent products, its hardware design has a high quality certification, this advantage also extends to the hardware design of binocular modules.

In view of the large amount of calculation of binocular cameras, DJI believes that five or six years ago, car companies might have spent a lot of effort to get binocular vision on the bus, but it did not work very well, in part because of the large amount of calculation. Today, the addition of some parallel computing can alleviate the problem of large amount of computation.

In addition, DJI has also done a lot of layout in binocular vision algorithm, which further improves the performance of the product.

As for the problem of binocular vision detection distance, DJI feels that this is mainly related to vehicle positioning. KiWi EV is only used in the city, and its driving environment is not complex, so it does not need to be detected too far away, only to ensure the accuracy of about 80 meters.

Later, if you need to detect a longer distance, the vehicle can upgrade a higher-pixel camera. At present, the KiWi EV only uses a 2-megapixel camera, and the 8-megapixel camera can be upgraded later.

What is more interesting is that DJI's scheme uses four fisheye lenses. In the designs of other car companies, fisheye lenses are mainly designed to detect the surrounding environment when parking, while DJI uses a unique algorithm to enable fisheye lenses to detect the side of the car, so that the side of the vehicle can be found early and the detection accuracy of the side of the car can be further improved.

▲ DJI can detect the side lane through fisheye lens.

03. Multiple scenes can be parked and automatically out of the parking space, and at the end of the test, the car also experienced the intelligent parking system in DJI.

DJI intelligent parking system has P5, P100 and P1000, which can assist drivers or completely autonomous valet parking in the speed range of 0~30km / h. The three solutions all support parking in indoor, outdoor, open or closed parking.

Among them, P1000 can realize autonomous valet parking, including long-distance autonomous parking and long-distance autonomous summoning. The driver only needs to park the car in a designated location outside the parking lot, get off and leave, and the vehicle will drive into the parking lot and find a suitable parking space.

When the owner needs to pick up the car, the vehicle will start on its own and drive to the driver.

The performance of DJI P100 intelligent parking system is slightly weaker, but the difference between DJI P100 and P1000 is only in the distance of remote control vehicles. P1000 can achieve long distance, while P100 can only achieve medium distance.

DJI P5 intelligent parking solution has parking assistance, depot assistance and other driving assistance functions.

From the perspective of specific functions, KiWi EV should use DJI P100 intelligent parking system. After pressing the parking button on the steering wheel, the vehicle begins to find parking spaces automatically and can automatically identify the parking line in the process of moving.

▲ Dajiang L2 parking button

After finding the parking space you want to park, select and click the start parking option, then you can leave with both hands and feet, and the vehicle can park automatically without requiring the driver to leave the vehicle.

Cars and things found a number of special scenes in the parking lot, and the vehicles finally finished parking.

In the case of a broken road, the parking space closest to the wall is generally very difficult to park, and KiWi EV can enter the parking space smoothly after adjustment. And the extremely narrow parking space, KiWi EV can also enter smoothly, at the same time, by pressing the parking button on the key, the vehicle can automatically drive out of the parking space, making it convenient for drivers to get on and off.

▲ Dajiang L2 parking in narrow parking space

For parallel parking, the completion effect of the vehicle is also good. The parallel parking space selected by the car is located on a half-slope road. After identifying the parking space, the vehicle finally parked smoothly on the parking space after many adjustments.

▲ DJI L2 parked in parallel parking spaces

The whole intelligent parking gives things a very good feeling, and the accuracy of finding parking spaces is extremely high, and after finding parking spaces, it can be parked smoothly, and even a handful of storage can be realized on some parking spaces.

However, at present, the system does not support memory parking, and when the OTA is updated, the function of reversing will be added.

04. Will deepen cooperation with Wuling firmly take a gradual route it is understood that this car in cooperation between DJI and SAIC GM Wuling is currently undergoing pre-test driving and is about to begin formal delivery.

The two sides have conducted in-depth cooperation in the process of developing the car, and many functions are defined according to the actual needs of consumers and the overall configuration performance of the vehicle.

For example, at the beginning, Wuling hopes that the car can go straight at the intersection and turn left and right. DJI believes that under the current computing power, it is not impossible to achieve such a function, but it is difficult to guarantee whether this function is safe in the actual landing.

In the process of turning left and right at the intersection, in addition to traffic participants such as people and bicycles, other vehicles will also be inserted or too close to the car, so it is easy to be dangerous to continue to use auxiliary driving.

Therefore, considering the security and usage scenarios, DJI finally developed only the function of going straight to follow the intersection.

In the process of cooperation, the two sides have been taking practicality as the first essence of the function. Wuling made a user emotion report, recording the user's mood in each scene, and then both sides defined some driving assistance functions according to the user's mood to ease the user's anxiety.

The joint development of the first car, the KiWi EV, has been going on for more than a year, and now it is finally the eve of delivery.

▲ KiWi EV

According to DJI, KiWi EV is only the first step in the cooperation between the two sides, which is still deepening, and more Wuling models may be equipped with DJI's intelligent driving technology in the future.

As for the evolution of autopilot, DJI does not recognize the current L1~L5 grading method, they believe that the main difference between L3\ L4 and L2 lies in the determination of responsibility, and there is no obvious change in this aspect.

DJI will also adopt a gradual approach to do it, gradually improve its functions, practice its basic skills, and then launch better products in the future.

05. Conclusion: the initial public information of Dajiang self-driving layout results shows that Dajiang only announced its entry into the smart car industry in 2021, but it began to layout as early as 2016, got a self-driving test license in 2018, and set up a car BU in 2019.

Therefore, the layout of DJI in the field of autopilot is not too late.

After the official announcement, Dajiang has carried out a large number of research and development and testing of intelligent driving assistance system this year, and has received an order from SAIC GM Wuling, which is only the first step, and the follow-up cooperation with Wuling will be deepened. Wuling, as the highest national car company, may also bring great promotion to Xinjiang.

On the other hand, DJI is also in contact with Volkswagen, Dongfeng and other car companies, once its intelligent driving system has been verified by the market, it may load more car companies.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.