Detection of Illegal Parking Events Using Spatial-Temporal Features

Abstract

In this work, we propose a novel deep learning framework, called Attention-Based 2-layer Bi-ConvLSTM (denoted as Att-2BiConvLSTM) model, to predict the number of illegal-parking events in urban spaces. We model the research as a "next frame" prediction problem, which aims to improve urban transportation conditions and enhance the security and right-of-way for pedestrians. Various features in the prediction model are considered: some of them (e.g., hourly weather, traffic volumes) are dynamic every hour, while others (e.g., road network, point-of-interests) are static. To boost the effectiveness of static features, we propose a dynamic training process to transform the static features into dynamics. After that, all features can vary with time so that they are capable of handling a real-time prediction scenario. Moreover, we propose an attention mechanism for enhancing our bi-directional ConvLSTM model. With experimental verifications, we find that our proposed Att-2BiConvLSTM model can outperform other state-of-art and baseline methods. Besides, our model is useful for combining all features to make an accurate prediction.