Automatic Quality Improvement of Data on the Evolution of 2D Regions

This work deals with data cleaning and quality improvement when representing the evolution of 2D regions extracted from real-world observations. It presents a method that combines quadtrees and time series to identify inconsistencies and poor-quality data in a sequence of 2D regions. Our algorithm splits a 2D space recursively into buckets and creates time series using spatial functions on bucket-delimited subregions. Smaller buckets represent the subregions with a higher number of inconsistencies over time. Then, it uses time series outlier detection methods and consistency metrics to identify polygons that are poor-quality representations and remove them from the original sequence. The proposed method identifies errors and inaccuracies even if they occur in several consecutive observations. We evaluated our strategy using a dataset extracted from real-world videos and compared it with another method from the literature. The results prove the effectiveness of this proposal.