Autonomous vehicles
Algorithms
As the hardware provides a continuous flow of data, it is the algorithms’ responsibility to interpret this stream of information into a live, digital representation of the world outside the vehicle. A self-driving car relies on hundreds of different algorithms, each with a specific role in the decision-making procedure. Some are responsible for creating a virtual landscape of the area around the vehicle by identifying surrounding objects, whilst other algorithms use these interpretations to calculate probabilities and predict movements of objects. Pattern recognition algorithms solely focus on object recognition. Advanced driver-assistance systems (ADAS) rely upon huge amounts of data. This data must first be filtered to recognize the relevant images containing a particular category of objects. The algorithm can filter objects by detecting their edges and other physical/structural characteristics such as circular arcs and line segments. This algorithm can also discard anomalies in the data and recognize patterns in the data before classifying the object. Naturally, for the algorithm to successfully categorize objects, an immense amount of machine learning and initial human input is needed. Essentially, programmers give these algorithms a range of similar images, and once the algorithm has sorted the images into various categories such as ‘traffic light’ or ‘road signs,’ the programmers give the computer feedback and identify where the algorithm has been correct or wrong. Over time, machine learning takes over and rapidly categorizes the images ( How Machine Learning Algorithms Made Self Driving Cars Possible? 2018). Millions of people have had to complete a ‘human verification test’ whilst using Google. When selecting the quadrants of images containing traffic lights, we are feeding the algorithm answers for it to learn from and replicate when used in an autonomous vehicle. This mass gathering of data from around the world allows the algorithm, and therefore its machine learning capabilities, to excel and evolve exponentially. Regression algorithms are used for object detection, object localization, as well as predicting movement. It uses repetitive or reoccurring aspects of an environment to create a statistical model of the relation between an image, and the position of a specific object within the image. Over time, the algorithm can learn about other objects (possibly using information provided by pattern recognition algorithms) and begin to recognize similarities without large amounts of user input ( How machine- learning algorithms 2018). Implied by its name, decision matrix algorithms provide information that is directly related to the final outcome/action of the vehicle; it is the physical ‘decision - making’ phase in the overall proc ess. This type of algorithm identifies, analyses, and rates performance of relationships between sets of values relative to the information that they were derived from. The most commonly used decision matrix algorithms are ‘Gradient Boosting’ (GDM) and ‘AdaBoosting’ . Decision matrix algorithms are made up of independently trained, multiple decision models, whose individual predictions are combined to create an overall possibility or outcome while minimizing uncertainty. The algorithms must correctly classify, recognize, and predict (the movement of) almost all objects in the area every millisecond; the accuracy of these algorithms is responsible for the final decision that the vehicle makes. By weighting the probabilities of different scenarios that the vehicle might find itself in at any given time, as well as the success rate of any action the vehicle might take in accordance with these scenarios (like evasive action or braking), the car can make a calculated decision and choose the outcome with what it deems as the ‘safest result’ . With the use of machine learning, it is possible that the vehicle’s computer logs any successful decisions it may have to make, so in an identical or similar circumstance, the vehicle can
19
Made with FlippingBook interactive PDF creator