Efficient City-Wide Multi-Class Multi-Movement Vehicle Counting: A Survey

Abstract

With the advent of accurate deep learning-based object detection methods, it is now possible to employ prevalent city-wide traffic and intersection cameras to derive actionable insights for improving traffic, road infrastructure, and transit. A crucial tool in signal timing planning is capturing accurate movement-and class-specific vehicle counts. To be useful in online intelligent transportation systems, methods designed for this task must not only be accurate in their counting, but should also be efficient. In this paper, we study the multi-class multi-movement vehicle counting problem, overview the state-of-the-art methods designed to solve this problem, and present a series of comprehensive experiments, using traffic footage with O(5) vehicles captured from 20 different vantage points and covering various lighting and weather conditions. Our survey aims to answer the question whether we are ready to leverage traffic cameras for real-time automatic vehicle counting. The results of our analysis show several promising approaches and identify areas where additional improvement is needed.

An Extreme-Adaptive Time Series Prediction Model Based on Probability-Enhanced LSTM Neural Networks

“…However, methods such as Autoregressive Integrated Moving Average (ARIMA) (Box and Jenkins 1976) seem to adjust poorly to extreme changes in the water level values and cannot easily find the nonlinear relationships among the data. Recently, deep neural networks (DNNs) have shown their great advantages in various areas (Yang et al 2019;Anastasiu et al 2020;Pei et al 2021). Both conventional Neural Network (NN) and Recurrent Neural Network (RNN) models have been used to overcome the disadvantages of traditional methods for time series forecasting (Qi et al 2019), since they can map time series data into latent representations by capturing the non-linear relationships of data in sequences.…”