Learning to Discover Curbside Parking Spaces from Vehicle Trajectories

The increasing number of vehicles makes it difficult to find parking spaces, especially for the citizens living in metropolises. Therefore, many people prefer parking at curbsides for convenience’s sake regardless of whether they are authorized or not for parking. It tends to cause a severe problem of traffic jams. In order to address the issue, urban planners are eager to mitigate the stress of the traffic by discovering more curbside parking spaces. The technique of image processing supports a modern solution that can only cover a small range of roads with the aid of CCTV surveillance systems. In this paper, we propose to leverage large-scale vehicle trajectories in Baidu Maps to help detect more popular curbside parking spaces, ensuring broad coverage of roads. However, the main challenge we encountered is that only a small proportion of curbside parking spaces are authorized/labeled. To tackle this challenge, we distill several effective features and develop an approach based on a weak-labeled active learning framework. The framework can utilize very few annotated training samples to label the rest unlabeled samples iteratively. In each round, our approach can adopt several pre-trained classifiers that are ensembled to find more positive or negative samples. We use two real-world datasets to measure the performance of several modern methods of curbside parking space detection. Extensive experimental results demonstrate that our method achieves significant improvements in Fl-score over the other approaches on the two datasets.