Browsing by Author "Costa, Pedro Miguel Marques"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Road intersection estimationPublication . Costa, Pedro Miguel Marques; Ramos, Carlos Fernando da SilvaIn this work, road intersection estimation problem is addressed. In autonomous driving, safety is one of the most important concerns. Intersections cannot be neglected. According to the United States department of transportation, around one-quarter of traffic fatalities and about one-half of all traffic injuries are attributed to intersections. Until now, many works detect it indirectly by detecting road markings, road signs, traffic lights and Advanced Driver Assistance Systems (ADAS) maps information. Intersections can be considered as a background object. Background has been oversight in most of the object detection researches that focus on foreground objects. The approach used, leverages camera and Light Detection and Ranging (LiDAR) sensors’ information to predict the existence and distance to an intersection, more specifically its entry and exit. In fact, intersection’s limits are subjective. The goal is the detection of a zone described as non-stopping area. Its entry is the most important information. It is defined as last safe stop location before the intersection and highly important to make a velocity adjustment of the vehicle in such a way that the car stop in a safe position while slowing down with comfort. A novel dataset was created due to no previous found dataset with the desired labels. Deep learning Convolution Neural Network (CNN) architectures are implemented making use of Bird’s Eye View (BEV) transformations. The proposed architecture comprise sensor fusion approaches, but tests were performed to evaluate the performance of each one independently. Results achieved proved that it is possible to detect intersections with the sensors and methods proposed. Both sensors’ backbones proved to be able to detect intersections. Anyway, LiDAR gets better performance than image camera approach. The fusion of both sensors revealed to be the best solution, taking advantage of both sensors’ complementary information. Simple intersections with good visibility achieved the best score while some complex intersections as the case of plural junctions, revealed to be more challenging.