Jaime Fernandez

Ph.D. Student

Machine Learning & Statistics, Media Analytics

Autonomous Vehicles like self driving cars are becoming a popular idea. Using them would avoid and reduce accidents on the roads. In order to become autonomous, a vehicle is equipped with sensors and a system called Autonomous Driver Assist System (ADAS). Currently, cameras are becoming popular as a type of sensor in autonomous vehicles, but as we know, while a camera records images, it’s a lot more complicated to make a camera “see” and interpret what it is “seeing.” s. We do this with help of techniques of image analysis and Artificial Intelligence. Using these techniques, semantic information, or information that tells the computer what is going on, can be obtained from images and videos. This semantic information consists of the objects that surround a vehicle, i.e. pedestrians, other cars, traffic lights, lanes, distance to these kinds of objects, etc. This information is then used to make decisions like when to brake e.g. if a person or a vehicle is in front of the driver.

In order for an autonomous vehicle to navigate its way around its environment, it needs appropriate models of its  static environment (ie the roads and roundabouts it’s on) and its dynamic environment (ie the moving objects and traffic around it). Remarkable progress has been made in moving-object-free scenarios. In contrast, environments with a large number of moving objects – in particular inner-city locations with many pedestrians and cars – still pose significant challenges. Vehicle and pedestrian detection also have made big progress over the years, and methods that allow the detection of these are more reliable nowadays. Detecting these kind of moving objects is already useful but predicting the location of these objects in the future is of great importance for several applications.

Predicting the path of other traffic participants is a crucial task for autonomous vehicles for safe and efficient driving.  It is also necessary for many ADAS where both the ego vehicle and other road users’ trajectories have to be predicted. Even if path prediction is not a deterministic task, it will be possible to identify the most likely position of a moving object in the future.