Abstract:The achievement of level 5 autonomous vehicles on urban roads requires performance equal to that of a human driver in every scenario. While many obstacles that stand between achieving this goal have already been overcome, many problems have yet to be solved. In this paper, we will address the specific challenges bicycles pose for self-driving cars in urban environments. One of the most prevalent obstacles is detection and tracking of bicycles. Their relatively transparent profile, ever changing as the bicycle moves, and their slight frames make detection a difficult problem. Furthermore, their ability to quickly maneuver in cluttered urban environments can generate inaccurate tracking models and faulty prediction estimates. Significant work has been done in sensor and algorithm development to solve the bicycle detection, tracking, and prediction problem, yet problems remain as datasets and algorithm analysis are not accessible to academic researchers. This information is instead considered proprietary. Of the published work in this field, most approaches use idealistic datasets that do not accurately represent real world conditions in order to improve the quality of their results.
To further the development of LiDAR sensors and algorithms this paper introduces the first open LiDAR dataset, collected in real world environments. The author presents realistic datasets taken with affordable sensors, along with qualitative performance results of leading algorithms.
Easy access to this dataset and analysis allows researchers and developers to create systems and algorithms that perform in real world scenarios.
Professor Hanumant Singh (Advisor)
Professor Taskin Padir
Professor Robert Platt