Learning on-street parking maps from position information of parked vehicles

Abstract

Many drives in crowded cities end with a challenging parking search, and visitors often do not know which streets allow on-street parking. Therefore, we present a learning-based approach to automatically generate on-street parking maps from parked vehicle positions detected by sensing vehicles. Multiple sets of features are proposed to describe the occupancy of every small road segment and its surroundings at different time instances. The usage of k-means algorithm as unsupervised learning and random forests as supervised learning are compared by applying these feature sets. The proposed approach is evaluated with repeated LiDAR measurements on more than five kilometers of potential parking space length. Our approaches, while keeping the model more generic, reveal slightly better results than an approach from literature. In particular, the unsupervised approach does not need a training data set and is free of any area specific parameter choice.