Sensors on AV

The following is a list of popular sensors on any autonomous vehicles (AV):
2D camera, range camera, LIDAR, Radar, Sonar, GPS, IMU/compass, odometry.

Since autonomous driving requires the vehicle to sense the environment as good and even better than human drivers, multiple sensors are usually used for compensate their capabilities.

To achieve autonomous driving, the vehicle need to sense the followings:
Location, heading, map, traffic signs, signal lights, surroundings: other vehicles, pedestrians, biker, motorists, obstacles, curb, poles, gates, fence, kids jumping in front the vehicle, etc.

The sensed information should be processed very quickly so that the vehicle can make a decision in several milliseconds (ms). To do so, usually the latency requirement for the sensors should be less than 2-3 ms.

The sensors should be not very expensive. However, there could be different cost ranges for different level of autonomy. For level 3 autonomy, since the vehicle only provide some comfort to the driver, but doesn’t generate any new revenue, to be competitive, the sensors should cost less than 10% of the cost of the total vehicle. However, for level 4 autonomy, the the vehicle could be shared and then used 24 hours a day, it does generate more revenue, the cost of the sensor could go up to 50% of the total vehicle.

Now, let’s looked each individual type of sensors and their pros and cons, so that we can understand why they are used.

2D camera:

Pros: High resolution, high speed, rich information, low cost.
Combine two 2D cameras, we can have stereo information that provide 3D information.

Cons: low dynamic range (saturate with high lights, sun), need a lot computation power, no 3D information for single camera, time to contact is not necessary reliable, 3D from stereo is not necessary reliable.

Because the pros and cons, cameras are good for provide details of the surroundings. The images are used for recognition (what the things are), detect and interpret traffic lights, signs, lanes, detect vehicles, pedestrians, bikers, etc.

Combining with other sensor, it is more powerful. For example, lane detection improves cars’ GPS self-localization estimation.

Israeli Company Mobileye

Ultrasonic sonar,
Sonar has an emitter that emit ultrasonic sound at 50 KHz, the has a receiver to listen to the bounce back of the sound. It measures the time of flight of the sound.

Cheap, simple, one signal. Relative reliable and fast.

Very low resolution, cone shape, sensitive to dirt, usually used for near range.

So, sonar sensor for AV are usually used to detect close objects and mainly for near obstacle detection. It usually works with auto break to stop the vehicle and parking to detect other vehicles and curb. For example, Google car has ultrasonic sensors on two rear wheels for parking and reverse. Because they are cheap, they could be placed at multiple locations of the vehicle.

Full name: LAser Detection And Ranging or LIght Detection And Ranging

Similar to ultrasound sonar, it also measure time of flight, but time of flight of a Laser beam. With one Laser emitter and receiver, a scanning mechanism is used to scan the Laser beam and measure the bounced Laser at each angle. The scanner is usually a rotating mirror that redirects the laser beam.

Different LIDAR devices have very different ranges. Some low cost ones had a-few-meter range and some high end has upto 200 meter. The scanning frequency ranges from 1Hz to 100 Hz or more.

For a scanning LIDAR, many of them has 1-degree resolution and some of them has sub-degree resolution.

The line scanner is usually less expensive. The price ranges from a few hundred dollars to several thousand dollars. The price is usually related to the resolution and distance.

A 3D LIDAR gives 360 degrees scan of surroundings. It usually has multiple laser beams and doing 3D scanning. It usually very expensive. The scanning speed is relative low, a few Hz. High end 3D LIDAR has 1-degree resolution and has 200 meter range. It usually obtains tremendous amount of data. For example, Velodyne’s existing HDL-32E sensor that scans at 700,000 points per second. Some models delivers 1.6 million 3D points per second. They are usually very expensive. The cost of making one is not high, but because they can only sell a small number of them each year, the unit price is very high. Velodyne sold a couple hundreds of them last year. The price can go down to a few hundred dollars if the demand increased to millions units a year.

Pros: reliable 3D points. Not sensitive to lighting. A lot of information

Cons: expensive, no color information, has problem with reflective surfaces and transparent surfaces. Need high computation power to process the information. The scanning speed is not very fast.

3d LIDAR is not widely used. They provide very reliable and rich 3D information of surroundings. Level 4 autonomy could equip 3D LIDAR. It is too expensive for Level 3 or under.

Full name: Radio Detection And Ranging
Similar to the two above, it emits radio signal (electromagnetic waves of the radio spectrum) and receives the bounced signal, and measure the time of flight. Doppler radar measures the shift in frequency in the bounced signal and computes the speed from the frequency change. So Radar can detect both distance and the moving speed (relative) of an obstacle. Doppler Radar along doesn’t respond to stationary objects.

Radar can detect multiple objects in a range of 30 -100 meters or more. High end Radar usually has a very long range. Radar is very good in difficult weather conditions, it cuts through fog rain and snow with no problem. It is not sensitive to dust. However, since the signal is electromagnetic wave, it doesn’t detect painted wood or plastic (stealthy to Radar, so airplanes have coating to not to reflect radar signals). Human is also almost transparent to Radar. Radar is very sensitive to metal surfaces. Curved metal surface could be perceived as a large surface. So a small aluminum can on the road could be detected as a large obstacle. It doesn’t work well on bridges and in tunnels.

Pros: Relative less expensive because of mass production. Detect other vehicles very well. Reliable to certain material, fast response, simple, robust to bad weather.

Cons: not responsive to some material. Difficult to tell the size of an obstacle. Relatively low resolution.

So, it is commonly used for automatic cruise control, for lane change assistants and in emergency brake system.

Google car has four Radar sensors. They are mounted on the bumpers — two front and two rear. They are used to see the car ahead and maintain distance. Tesla car relies more on Radar sensors than cameras now.

There are actual efforts to improve the resolution of Radar using mm-Wave radar signals.

GPS, Odometry, and IMU.

GPS is not accurate and not always available. It provides rough positon information and available 99% of the time.

Odometry is not accurate and only one ready.

IMU measures acceleration. High-end and military grade IMUs are much more accurate. High accuracy IMU is a research topic.

In summary, there is no silver bullet that can solve everything. AV needs multiple sensors. Signals from multiple sensor should be fused together to provide rich and reliable information. For example, Google car has one Velodyne 360 LIDAR (64 laser beams, 200m range), a front camera for near vision, four Radar sensors mounted on the bumpers (two front and two rear), GPS, IMU, Ultrasonic sensor on two rear wheels.,

Leave a Reply

Your email address will not be published. Required fields are marked *