cool hit counter The Intelligent Eye to the Future of Autonomous Driving_Intefrankly

The Intelligent Eye to the Future of Autonomous Driving


learned

Bole

wise

winnow

special column

12

Experts in this issue.

Feng Hao

Director of Marketing and Strategic Development, Bosch Chassis Control Systems China. Dual degree in Germanic Languages and Literature from Beijing Foreign Studies University and Business Management from the University of Worms, Germany. He joined Bosch in 2012 and has held the positions of Active Safety Product Manager, Marketing Communications Manager and Senior Product Manager for Driver Assistance.

In the last op-ed, I'm sure you've had a first look at driver assistance systems, and the driver assistance features that can be achieved by relying on a little radar.

Click here for a refresher

Today, let's take another look at who provides a pair of eyes for vehicles on the road to the future of autonomous driving?

Video cameras and the functions performed through them

A video camera, or "camera" for short, as the name suggests, is an image sensor that processes images to enable object detection and then different driver assistance functions.

The camera is similar in structure to the camera in our mobile phones or DSLR cameras, in that the light entering the lens is reflected on the CMOS (Complementary Metal Oxide) imaging sensor through an optical component in front, which produces a frame by frame digital image, which is then analyzed by the image information processing unit (ISP) built into the camera to detect vehicles and their lights, pedestrians, lane lines and other targets in the picture.

According to the number of optical components, front-facing cameras for vehicles are subdivided into monocular cameras, binocular cameras, and tricameras, which are used in models such as Tesla.

Monocular camera With only one optical lens, the algorithm of the famous Mobileye company is based mainly on a monocular camera for object detection.

As we mentioned before, radar detects objects by means of echo reflection, a direct approach that is very accurate for detecting longitudinal motion information (e.g., distance, etc.) of objects. If we compare a monocular camera to one of our eyes, all it collects is a two-dimensional picture that lacks distance. So how does it restore the distance of the object as well as the speed of its movement?

Algorithms are critical. In the case of Bosch's next-generation monocular camera, for example, one of the core algorithms it has is to reduce this information to depth information by comparing the intensity change in the amount of light received at the same pixel point in the two frames before and after. With the help of an algorithm called "structure from motion (SfM)", the distance and relative speed of the detected object in front of the car can be deduced, achieving the same effect as radar.

As the algorithms continue to be optimized, the monocular camera is as accurate as radar at certain ranges. However, as it is a projection, its accuracy cannot be fully reliable under complex weather conditions. This is also a certain "inherent deficiency" of the monocular camera.

By adding another optical lens and spacing the two lenses at a certain distance apart, it becomes Binocular camera

The binocular camera works on the same principle as the human eye, by comparing the same image taken by both lenses at the same moment and comparing the distance difference between the same pixels on different images (the small difference in response at the pixel level between the two lenses due to the apparent distance) to determine the distance, i.e. 3D parallax. This is also a direct type of measurement, and unlike the monocular camera ranging principle, the binocular camera based algorithm produces a more reliable target than the monocular.

As an example, for example, when starting, if the car does not move and the car in front of it does not move, due to the nature of the SfM algorithm requires movement in the picture, then the distance between the car in front and the car at this time is not known by the monocular camera, while the binocular camera can easily solve this problem.

Using the binocular camera, most driver assistance functions other than ACC can be achieved without the need for radar. Another use is to calculate the space available for the vehicle to travel in front of by using the 3D parallax, which we call "free space". Free space is an essential piece of information to be applied in an autonomous driving algorithm.

Of course, with great functionality comes increased costs. Currently, binocular cameras are mostly used in some high-end cars.

While the application of Tesla's Tri-eye camera In the case of the Fisheye camera, three monocular cameras with different detection distances, different fields of view and different functions are combined, including a telephoto camera with a narrow field of view but a long detection distance, a standard focal length camera with a wide field of view but a moderate detection distance, and a fisheye camera with an opening angle of 180° or more for proximal detection. We will not go into details here.

For front-view cameras, whether monocular, binocular or trinocular, their main purpose is to detect objects in the traffic ahead, and depending on the type of target detected, the functions they support are divided into the following categories.

- Based on the detection of vehicles, pedestrians, and cyclists, etc. Safety features such as Forward Collision Warning FCW, Automatic Emergency Braking AEB, etc. can be implemented. That's right! The same function can be achieved with different types of sensors, for example, the AEB function is available in mass production with both radar and camera.

- Based on the detection of lane lines, the The LDW lane departure warning function, which warns the driver of unintentional lane departures, and the RDP lane correction system, which intervenes in lane departures in conjunction with the electric power steering EPS system, are possible, as is the LKA lane keeping centering function, a comfort function that keeps the vehicle within a certain area in the center of the lane by detecting the lane lines on both sides.

This year, Euro NCAP, the European New Car Assessment Programme, evaluated the emergency lane keeping function as a mandatory on item, meaning that if the driver is judged to be drifting towards the edge of the road with the lane keeping system switched off, the system will also switch on urgently to intervene and help the vehicle get back into its lane.

- Based on the detection of traffic signs, the Traffic sign recognition TSR function that can provide information to the driver such as speed limit signs, no overtaking signs, etc. In conjunction with the cruise control system or speed limit system, the intelligent cruise function iACC, etc. is also available.

- Based on the detection of lights from oncoming traffic. It can provide control of the car's lights, such as the simplest automatic high and low beam switching HMA function, which prevents the car from causing dazzling to the driver of the opposite car when the high beam is on; when combined with the matrix headlights with adjustable partial beam, it can also turn down the high beam in some areas ahead. This feature, called Adaptive Headlights ADB, is very useful in China to maintain the view ahead of our car while preventing dazzling to the driver of the opposite car.

The features we have talked about above are based on the front view camera to achieve, in the camera field there is also a ring view camera, its optics can provide a field of view from 130° to more than 180°, the pixels range from 300,000 to 1 million, can provide a close range image of about 10 meters, because its physical optical module provides an image very much like the image seen by the eyes of a fish, so it is also called Fisheye camera(fish-eye camera)。

These cameras are generally used for parking assistance related functions and can provide a rear view image, commonly known as a reverse camera. These cameras can improve the driver's visibility around the vehicle during parking, preventing collisions with pedestrians or obstacles and greatly improving safety. In 2014, the U.S. has made back-over avoidance (Back-over avoidance) mandatory as standard through legislation, i.e., there are mandatory requirements for rear vision , which, of course, is also related to the national conditions of the U.S. SUVs and pickup trucks are predominant.

A more advanced 360° surround view system as well as advanced park assist and auto park functions can be supported by the fusion of four surround view cameras. It is through this multi-camera fusion that Bosch's full 3D 360° Surround View system, offered on the new BMW 7 Series in 2016, is made possible.

This system enables a better transition effect and experience through its own graphics processing acceleration at the hardware level (similar to GPU graphics acceleration), in addition to standing up traditional flat images for 3D surround view visuals.

At the same time, the most difficult to deal with distortion and distortion when stitching images, also through hardware and software processing to achieve better results compared to other similar products.

The sensors used in driver assistance systems (e.g. radar and cameras) each have their own advantages and, of course, their own limitations. For example, radar is insensitive to the shape of objects, and camera detection performance is reduced in low light and poor weather conditions ...... This requires us to apply a combination of different types of sensors to complement each other's strengths and weaknesses.

By fusing different types of sensors and their data, more sophisticated driver assistance functions and even autonomous driving can be achieved. In the next issue, we'll talk about how Bosch combines and optimizes these different types of sensors in order to make them work.


Recommended>>
1、The Chinese AI market is so fat that Google still cant let go without stopping
2、Waller教授团队Nature发文深度神经网络和符号人工智能助力化学合成
3、SuperBook Project Research NotesFabric Network Node Architecture
4、How much energy a business consumes can lit check it out once on the net
5、澳洲AI研究生专业解析

    已推荐到看一看 和朋友分享想法
    最多200字,当前共 发送

    已发送

    朋友将在看一看看到

    确定
    分享你的想法...
    取消

    分享想法到看一看

    确定
    最多200字,当前共

    发送中

    网络异常,请稍后重试

    微信扫一扫
    关注该公众号