In the context of self-driving cars, it is an imperative that the vehicles can securely communicate with each other. As current technologies cannot guarantee this (without the use of a public key authority), the idea of using an out-of-band channel was conceived. For a usable out-of-band channel, that can be accessed without installing additional hardware in vehicles, braking lights were chosen. The sent messages will be received on already existing cameras in the car, such as the cameras used by a lane keeping assistant. In the scope of this project, the following tasks should be accomplished:
(1) Recording of braking light images
To simulate data transmission via braking lights in a real scenario, it is necessary to have access to test pictures. These pictures should be taken with actual lane keeping cameras, and depict the leading car transmitting messages on its braking lights in a variety of scenarios. Possible scenarios may include different weather conditions (rain, snow, fog, clear), different lighting conditions (daytime, night-time, twilight), or different traffic situations (traffic jam, light and heavy traffic), and any combination of these. The goal is to compile a database of images, that show a number of different cars in all scenarios, with the additional information of (1) which car is seen, (2) the speed the recording car was moving at, and (3) the distance to the vehicle of interest.
(2) Automated annotation of braking light images
Due to there being many different conceivable scenarios in which we want to transmit data via braking lights, we need to opt for different braking light state detection algorithms. To be able to decide which algorithm to use, the system must be able to automatically assess in which scenario it currently is. In addition to automatically determining the situational context, a division of the image into regions (e.g. „road“, „buildings“, „sky“, „oncoming traffic“) and, most importantly, detecting the leading car, is also desired. The goal is to create a classifier (using a machine learning approach) that is able to swiftly and correctly add tags to the image and detected regions in the image, respectively.
(3) Automated selection of detection algorithm
By using annotated images of braking lights, we can simulate a multitude of scenarios, and test different braking light detection algorithms. Based on the data that is obtained in the simulation, we can decide which detection algorithm works best in which scenario. The goal of this master thesis is to create a tool (using machine learning approaches) that will automatically select the best-suited detection algorithm based on the situation information in an annotated image.