Vehicle Detection System for Smart Crosswalks Using Sensors and Machine Learning

Abstract

Cities are transforming into smart areas thanks to several key technologies involving artificial intelligence (AI), 5G or big data aimed at improving the lives of their inhabitants with new services (e.g., transport systems, including road safety). In this field, the paper describes how to improve vehicle detection through several machine learning techniques applied to smart crosswalks. As a main advantage, this approach avoids readjusting labels in classic fuzzy classifiers that typically depends on the system location and road conditions. To address this, various AI methods were evaluated with data taken from real traffic pertaining to roads in Spain and Portugal. The machine learning techniques were random forest (RF), extremely randomized trees (extra-tree), deep reinforcement learning (DRL), time series forecasting (TSF), multi-layer perceptron (MLP), k-nearest neighbor (KNN) and logistic regression (LR). The results were validated through a receiver operating characteristic (ROC) analysis, obtaining the best performance in RF with a true positive rate (TPR) of 96.82%, false positive rate (FPR) of 1.73% and accuracy (ACC) of 97.85%. This was followed by DRL and TSF, while MLP and LR presented the worst outcomes.