Advantages and disadvantages of using automotive vision systems in ADAS and automated driving

Lei Feng Network (search "Lei Feng Network" public concern) : Author of this article Zhu Yulong, Automotive Engineer. Mr. Zhu shares an industry article every day. Although the content is short, it is a kind of sharing after careful discussion. He hopes to help car practitioners.

Following the article “Let's Talk about Fire Extinguishing of Electric Vehicles”, this article mainly discusses the current ADAS and future autopilot, using the vision system as the strengths and weaknesses of the environment entrance, and the difficulties that need to be overcome.

The visual system is used as a necessary means in lane recognition LDW, signal recognition TSR, and pedestrian recognition. The pedestrian recognition in AEB will put forward higher requirements on the visual system.

Remarks:

Lane line recognition and centering complements high-precision maps and positioning

TSR can complement the communication with V2I

Pedestrian recognition is really difficult, the cost of multi-line Lidar is too high

The core strengths of the visual system are the distinction between "human" and objects, which is the distinction of priorities.

Vision system resolution is much higher than other types of sensors => Provide details of the road environment => Create a complete environment model

Vision System recognizes the shape and appearance of an object = Read external information

Focusing on V2I here, theoretically, in a complete V2X world, it is entirely possible to perfectly complement each other with interconnectivity and vision systems.

Today's exchange of laps down, the nets are inevitably idealized, and the perfect fusion of the two requires a lot of effort. You don't know so much if you don't do it.

Of course, there is a large visual killer, based on the video stream, through the look around to establish the external environment model.

1) E2E DNN

2) Stratified DNN learning

This direction requires a lot of computing power, learning through all levels, from the new driver to the old driver...

In the early stage of the industry, establishing a complete driving behavior learning mechanism also requires an internal visual system to record and compare.

Face look no distractions need to record

Optional: The steering wheel operation needs to be recorded (this can also be recorded by the sensor)

Optional: Throttle and brake operation (this can also be recorded by sensor)

In essence, this is actually a learning mode. After the model is established, it is contrasted and studied.

In summary, the problems faced by the system include: Some of the following problems can be solved, some can be improved, some are not, and external basics can be classified.

1, more interference and restrictions

1) Ghosts at Tunnel Portals and Tunnels

2) Frontal glare

3) Drops of water in front of the Camera

Windshield fog, snow, dust, dust or frost,

Dew or dust in the windshield

4) Various strange models

2, the impact of the environment is relatively large

Bad weather (such as heavy rain, snowstorm or dense fog)

Poor visibility, haze, smoke or water vapor

Rainy road vehicle reflection

Ambient illumination is relatively low (night & tunnel) + no car in front

3. Situation of external environment vehicles and lane lines and signs

Vehicle tilt and different angle scenes

Diffuse lane lines, invisible lane lines under rain and snow may not be recognized

Cover snow in front of the car, making the vehicle uniform in color

The slope of the road is relatively steep

Other interesting things are the following two cases.

This image is reversed.

This comes from outside the system, the system uses the brake light switch to collect people involved in the brake, but the failure of the switch automatically shut down the system

After the failure of the camera and camera processing chip inside the camera, look for information.

1) The hardware and software of the vision system is improving quickly, and many things need to be continuously updated

2) V2X I think we can just take it out and compare it. Tomorrow we will take a closer look and see how these two things are viewed in the system.

Lei Fengwang Note: This article updates the article for the author series, reproduced please contact the authorize and indicate the source and author, may not delete the content.