Predicting on-street parking violation rate using deep residual neural networks

Abstract

• Deep learning-based on-street parking violation prediction system is proposed. • The proposed method works in an indirect way, i.e., without using sensors. • A data augmentation and smoothing method is used to deal with noisy data. • Evaluated using a dataset with 3.9 million scans for illegally parked cars. The lack of available parking spaces can be among the most significant issues that can affect the quality of life of citizens in large cities. This has led to the development of on-street parking systems that typically ensure that parking spaces will be available for the local population, as well as provide easy access to parking for visitors, e.g., by providing directions for finding sectors where parking slots are available. Unfortunately, such systems are affected by illegal parking, i.e., parking without paying the parking fee, since in this case, the number of registered parked cars does not match the number of cars that are actually parked, leading to providing incorrect suggestions to drivers. This can discourage drivers from using such systems, potentially further increasing the parking violation rate. Such phenomena can be addressed by using smart sensors that can detect the presence of cars in various areas. However, installing and maintaining such systems is costly, which usually discourage cities from implementing such solutions. The main contribution of this paper is a Deep Learning (DL)-based pipeline that works in an indirect way (i.e., without using sensors) and allows for developing an accurate fine-grained parking violation prediction system, increasing in this way the accuracy of the information provided to on-street parking systems with minimal cost. To deal with missing and noisy data we also propose a data augmentation and smoothing technique that can further improve the accuracy of DL models , when used in such scenarios. The effectiveness of the developed system is validated using experiments on a large-scale dataset, which contains more than 3.9 million scans for illegally parked cars collected by the municipal police in Thessaloniki, Greece.