Computer Vision News - February 2019

● Intelligent Loop-Closure Detection: Loop-Closure Detection recognizes the car has returned to a previously visited location. Current state of the art approaches use image processing to detect Loop-Closure, these are computationally intensive and very dependent on image quality and features. Application challenges: ● Map Self-Correction: The scene is constantly changing and needs to be constantly adjusted using mapping and localization methods. ● Local computation power: Map acquisition using the car’s embedded processor without access to cloud infrastructure is challenging, especially so for CNN methods, which have heavy computational requirements. ● Unique signature for large-scale regions: Automated Driving maps are very large and similar objects are viewed with high frequency. This requires disambiguation, which is usually solved through semantic interpretation or global reference. Automated Driving is a rapidly advancing field with a complex structure (see figure below). The application of Deep Learning to the field has also seen significant advances, but not yet achieved the performance required for production. Two main paradigms exist: 1) The two-stage approach -- which first constructs a semantic understanding (mapping) of the world, and then, at the second stage, makes driving decisions based on its semantic understanding. 2) The end-to-end approach -- which learns the driving decisions in a single stage. Computer Vision News Research 5 Research Computer Vision News