Building an Autonomous car

1. ROS framework and CAN connection

How can I talk to my car? The first objective of implementing an autonomous car is: "How can I tell the car to move according to my intentions?". Because we can tell that a car is also a robot, we can utilize ROS framework to make connection between the car and the computer. For getting used to ROS, visit ros wiki.

The link is directed to ROS classic wiki page. I used ROS classic, but actually ROS2 is improved version. Also, in case of you are using python3 for development, you surely need to use ROS2 for compatiblity.

We installed an industrial-grade Linux PC in the trunk of the car, connected the vehicle and the computer through CAN communication, and exchanged signals between the car and user application as ROS topics. Using ROS, we gained a route to communicate with car. So here, we were able to subscribe information of car ... such as steer, velocity, etc ... and send command to the car ... such as break=10, steer=-10 ...

Run roscore, run your programs which subscribes & publishes, check if your rosnodes are set properly by your design, and jump into coding the car's brain!

2. Making my car drive at 100kph.

I was in charge of developing smart cruise control(SCC) and autonomous emergency braking(AEB) system. SCC is all about "moving" the car "back and forth". The user application gets the user input such as "100kph". Here we have the desired velocity, but driving the car would not work by simply telling the car "run at 100kph!". Instead, we need to calcuate break and accel pressure to match the desired velocity.

PID control is simple but powerful way to staisfy requirements. Yes, finding kp, kd, ki for pid control might go through heuristic methodology, yet it is the most efficient way to get desired system outputs by controlling unlinearly related inputs, such as motor power, break pressure, etc.

3. Object recognition: Point cloud or Vision?

In AEB and SCC system, there's more than just controlling velocity of the car: The car needs to "see" and "understand" the objects.
It is obvious why a car needs to be able to perceive objects. Without object perception, the car will not be able to stop because it does not know that an object exists. Two processes are required to enable the car to "see": detecting the object from raw data and running an algorithm to make the car react to the current situation.

Recognizing the object

Our team utilzed LiDAR sensor and Mobileye for sensing object.
We used the source code of the AutoLidarPerception library to extract objects from the point cloud of LIDAR. For Mobileye's vision data, Mobileye separates objects such as vehicles and pedestrians and provides information via CAN communication. Point cloud data and vision data have their own advantages and disadvantages. Point cloud data can sense distance information well, while vision data can better identify what objects are. In addition, vision data may be slower to process than point cloud data.

Our autonomous driving system cannot afford to miscalculate distances and requires real-time data processing. Even a slight delay of 0.5 seconds could lead to a serious accident. Therefore, it may seem that LIDAR is more suitable, but it is difficult to accurately identify the type of object with point cloud data. Through the AutoLidarPerception library, we simply grouped the point cloud data into boxes and heuristically determined whether the object was dangerous or not based on the size of the box. This could lead to malfunctions, such as sudden emergency stops caused by incorrect recognition.

Therefore, I solved the problem by fusing vision data and LIDAR data. I used the object class information obtained from the vision data and the distance and position information from the LIDAR to determine the object's danger level and location. In addition, if data was lost in either sensor, I ensured that the two sensors assisted each other based on priority to avoid any data gaps.

Reacting to objects' existence

Using the approximation of the bicycle model with freedom degree 2 and experimental data, we calculated the range of the trajectory and stopping distance of the car depending on its current speed, acceleration, and direction angle. Based on the objects in the corresponding range, whether they are preceding vehicles or other objects such as mainly corner walls or safety cones, we executed speed control algorithms.