Deep Learning for Object Relationships: Applications to Road Safety and Bin Picking
Abstract
Two projects on the topic of object relationships will be presented during this seminar:
1. Distances between vehicles: In order to improve road traffic safety, there is interest in first identifying road segments and intersections where there is a high risk of accidents. Identifying these areas can be done by referring to accident statistics; however, a more proactive identification of high-risk road segments and intersections could be done by collecting other risk metrics before the occurrence of many accidents. As such, this project, in collaboration with Miovision, is on the topic of predicting distances between vehicles in order to identify close calls, which are presumably much more common than accidents, from wide-angle footage of road traffic. This seminar will present work done toward the creation of a dataset and a proposed architecture for predicting distances between vehicles.
2. Relative arrangement of two objects: In the context of the bin picking problem, it is desirable for a bin-picking system (robotic arm, sensor(s), and model) to analyze the scene and pick objects quickly. This can be particularly challenging in cluttered scenes where (i) the target object in the scene is partially occluded by other objects, (ii) non-target objects might physically interfere with the robotic arm and cause a grasp attempt to fail, (iii) grasping the target object might cause other objects that were physically supported by it to be moved and to potentially fall out of the bin. A model to predict object relationships was developed for use on the MetaGraspNet dataset, which presents at least three distinct difficulties: reciprocally occluding objects, occluded relationship boundaries, and small overlap areas. A graph-network model with the aim of addressing these difficulties will be presented.
Presenter
Auguste Lawrence Whelan Koh, MASc candidate in Systems Design Engineering