Kindle Notes & Highlights
The concept of a self-driving vehicle may seem like a purely modern invention, but a sketch made by Leonardo da Vinci over 500 years ago suggests otherwise. The drawing shows a self-propelled cart powered by coiled springs and featuring programmable steering based on an arrangement of wooden pegs.
Reaching SAE Level 5 would make roads safer for everyone. Autonomous vehicles do not get tired or distracted, they don‘t drink alcohol, and they don‘t violate traffic regulations.
Passive sensors work by absorbing ambient energy, while active sensors emit some form of energy into the environment and receive and measure the reflected signals. For example, cameras are passive, while radar and lidar are active.
Sensors that measure a vehicle’s internal state are called proprioceptive sensors, while those that allow the vehicle to ‘see‘ outside are called exteroceptive sensors.
There are two basic types of radar: echo and Doppler. Echo radars work as described above. By obtaining data from two or more echo radars placed in different positions, it is also possible to capture additional information on an object’s position, such as its angle. A Doppler radar further enhances this capability by analyzing wave phases. It does this by keeping track of each particular wave and detecting differences in the position, shape, and form of the wave when it returns.
Ultrasonic refers to acoustic waves beyond the human audible range, i.e., frequencies above 20 kHz. As their name implies, ultrasonic sensors use these high frequency acoustic waves for object detection and ranging. Animals, such as bats, use similar principles to detect and locate their prey in low-light environments.
Ultrasound performs well in most situations, since it is not affected by the majority of weather conditions. Although ultrasonic sensors provide relatively little detail, they are not dependent on light, which can be an advantage in cases where not enough or too much light could yield misleading results. They also work in rain, fog, and snow, as long as the sensor itself is not covered by dirt, snow, or ice. In indoor, city, or crowded environments, their ability to detect non-metallic materials offers an additional safety measure in regard to pedestrians.
In some countries, cameras are taking on an even more significant role in the auto sector. In the U.S., for example, rear-view cameras are mandatory in all new vehicles built from May 2018 onwards.
One of the biggest drawbacks of GNSS-based positioning is that it requires a clear line of sight between the receiver and the satellites to work reliably. That means GNSS-based positioning works best in open areas with unobstructed views and does not work at all in indoor areas such as garages and tunnels.
Optical gyroscopes are typically laser-based, and utilize a physical phenomenon called the Sagnac effect [14]. When two optical beams propagate in opposite directions in a rotating ring path, their propagation time, i.e., the time for the beam to return to its starting point, will differ fractionally; the beam that travels in the same direction as the rotation will take longer to return to the starting point than its counterpart, as shown in Figure 2.13. The angular velocity applied to the ring can be deduced by measuring the phase difference of the beams.
Once the computing platform understands what is happening in the vehicle’s environment, it can make decisions and send instructions to the actuator interface to perform whatever movements are required. This sense-decide-act loop lies at the heart of automated driving.
A few milliseconds of delay could have serious consequences, so near-zero latency is imperative. Take emergency braking, for example. At 100 km/h, a vehicle travels 28 meters every second. So if an SDV reacts one second too late while traveling at 100 km/h, the total braking distance is 28 meters longer than it should have been.
With so many types of sensors available nowadays, it is likely that the vehicle’s computing platform will need to support more than one network system, for example Controller Area Network (CAN) and Ethernet. CAN is the most common type of automotive network bus architecture. It has been employed since the late 1980s to support reliable message exchange between multiple vehicle controllers.
In general, perception in dynamic environments, i.e., environments with moving objects, can be decomposed into two major sub-functions: Simultaneous Localization And Mapping (SLAM) and Detection and Tracking of Moving Objects (DATMO).
both the observed map features and the vehicle’s reported location suffer from errors, and these errors increase the further the vehicle travels from its last known location. However, thanks to loop closure, once the whole route has been driven in a loop more than once, SLAM can generate a consistent and accurate map.
Kalman filters are a class of Bayesian filters that make the assumption that all noise in the system is Gaussian.
common sensors such as cameras, lidars, ultrasound, and radars use a polar coordinate system (angle and distance),
However, SLAM, as with other non-linear least-squares problems, generally does not have any closed-form solutions [25]. Therefore, solving the problem typically requires an algorithm that starts with an initial value, either randomly selected, guessed or heuristics-based, and iteratively minimizes the cost function until convergence. Some popular standard solvers are the Gradient Descent (GD), Gauss-Newton (GN) and Levenberg-Marquardt (LM) algorithms.
A manifold is a topological space that locally resembles Euclidean space, but globally might not be Euclidean
Support Vector Machine (SVM) [6] is one of the most popular and efficient algorithms for classification, which aims to find a separating hyperplane that optimally separates sets of different class labels. In most cases, it might be impossible to separate the classes using a merely linear function. However, the non-separable data might be linearly separable in a high-dimensional space, and an optimal separating hyperplane can be determined. Thus, with the help of some non-linear mapping (or kernel) functions, the input data is first transformed into a high-dimensional feature space, and
...more
Despite SVMs falling into irrelevancy this is one of the clearest explanations of how they work I’ve seen
In robotics literature, the problem of motion planning is generally divided into two sub-problems: path planning and trajectory planning. Path planning is the task of finding the shortest collision-free geometric path from the start point to the destination. Trajectory planning is the task of determining the sequence of motions, as a function of time, to achieve a smooth drive along the desired path.
