Predicting station-level hourly demand in a large-scale bike-sharing network: A graph convolutional neural network approach

This study proposes a novel Graph Convolutional Neural Network with Data-driven Graph Filter (GCNN-DDGF) model that can learn hidden heterogeneous pairwise correlations between stations to predict station-level hourly demand in a large-scale bike-sharing network. Two architectures of the GCNN-DDGF model are explored; GCNN_reg-DDGF is a regular GCNN-DDGF model which contains the convolution and feedforward blocks, and GCNN_rec-DDGF additionally contains a recurrent block from the Long Short-term Memory neural network architecture to capture temporal dependencies in the bike-sharing demand series. Furthermore, four types of GCNN models are proposed whose adjacency matrices are based on various bike-sharing system data, including Spatial Distance matrix (SD), Demand matrix (DE), Average Trip Duration matrix (ATD), and Demand Correlation matrix (DC). These six types of GCNN models and seven other benchmark models are built and compared on a Citi Bike dataset from New York City which includes 272 stations and over 28 million transactions from 2013 to 2016. Results show that the GCNN_rec-DDGF performs the best in terms of the Root Mean Square Error, the Mean Absolute Error and the coefficient of determination (R²), followed by the GCNN_reg-DDGF. They outperform the other models. Through a more detailed graph network analysis based on the learned DDGF, insights are obtained on the “black box” of the GCNN-DDGF model. It is found to capture some information similar to details embedded in the SD, DE and DC matrices. More importantly, it also uncovers hidden heterogeneous pairwise correlations between stations that are not revealed by any of those matrices.     

The station-free sharing bike demand forecasting with a deep learning approach and large-scale datasets

The station-free sharing bike is a new sharing traffic mode that has been deployed in a large scale in China in the early 2017. Without docking stations, this system allows the sharing bike to be parked in any proper places. This study aimed to develop a dynamic demand forecasting model for station-free bike sharing using the deep learning approach. The spatial and temporal analyses were first conducted to investigate the mobility pattern of the station-free bike sharing. The result indicates the imbalanced spatial and temporal demand of bike sharing trips. The long short-term memory neural networks (LSTM NNs) were then developed to predict the bike sharing trip production and attraction at TAZ for different time intervals, including the 10-min, 15-min, 20-min and 30-min intervals. The validation results suggested that the developed LSTM NNs have reasonable good prediction accuracy in trip productions and attractions for different time intervals. The statistical models and recently developed machine learning methods were also developed to benchmark the LSTM NN. The comparison results suggested that the LSTM NNs provide better prediction accuracy than both conventional statistical models and advanced machine learning methods for different time intervals. The developed LSTM NNs can be used to predict the gap between the inflow and outflow of the sharing bike trips at a TAZ, which provide useful information for rebalancing the sharing bike in the system.     

Taxi vacancy duration: a regression analysis

Taxi vacancy duration is a major efficiency measure for taxi services. A clear understanding of the various factors and their effect on vacancy duration is necessary for the optimal operational management of taxis. Previous research has only dealt with vacancy duration by assuming probability distributions and has not investigated heterogeneity in the data caused by various factors. We develop a parametric duration model using not only new operational characteristics but also variables associated with taxi demand, such as weather, land use, demographics, socioeconomic variables, and accessibility of public transportation. The model is applied to a large-scale New York City (NYC) taxi trip dataset that covers operations for 2013. The results show that all the attributes have significant associations with vacancy duration that follows a log-normal distribution. Our study is expected to help improve the efficiency of taxi operations by decreasing the time spent in vacant states.     

Driving analytics using smartphones: Algorithms,comparisons and challenges

The present work investigates the use of smartphones as an alternative to gather data for driving behavior analysis. The proposed approach incorporates i. a device reorientation algorithm, which leverages gyroscope, accelerometer and GPS information, to correct the raw accelerometer data, and ii. a machine-learning framework based on rough set theory to identify rules and detect critical patterns solely based on the corrected accelerometer data. To evaluate the proposed framework, a series of driving experiments are conducted in both controlled and “free-driving” conditions. In all experiments, the smartphone can be freely positioned inside the subject vehicle. Findings indicate that the smartphone-based algorithms may accurately detect four distinct patterns (braking, acceleration, left cornering and right cornering) with an average accuracy comparable to other popular detection approaches based on data collected using a fixed position device.     

A novel passenger flow prediction model using deep learning methods

Currently, deep learning has been successfully applied in many fields and achieved amazing results. Meanwhile, big data has revolutionized the transportation industry over the past several years. These two hot topics have inspired us to reconsider the traditional issue of passenger flow prediction. As a special structure of deep neural network (DNN), an autoencoder can deeply and abstractly extract the nonlinear features embedded in the input without any labels. By exploiting its remarkable capabilities, a novel hourly passenger flow prediction model using deep learning methods is proposed in this paper. Temporal features including the day of a week, the hour of a day, and holidays, the scenario features including inbound and outbound, and tickets and cards, and the passenger flow features including the previous average passenger flow and real-time passenger flow, are defined as the input features. These features are combined and trained as different stacked autoencoders (SAE) in the first stage. Then, the pre-trained SAE are further used to initialize the supervised DNN with the real-time passenger flow as the label data in the second stage. The hybrid model (SAE-DNN) is applied and evaluated with a case study of passenger flow prediction for four bus rapid transit (BRT) stations of Xiamen in the third stage. The experimental results show that the proposed method has the capability to provide a more accurate and universal passenger flow prediction model for different BRT stations with different passenger flow profiles.     

Adaptive traffic signal control with actor-critic methods in a real-world traffic network with different traffic disruption events

The transportation demand is rapidly growing in metropolises, resulting in chronic traffic congestions in dense downtown areas. Adaptive traffic signal control as the principle part of intelligent transportation systems has a primary role to effectively reduce traffic congestion by making a real-time adaptation in response to the changing traffic network dynamics. Reinforcement learning (RL) is an effective approach in machine learning that has been applied for designing adaptive traffic signal controllers. One of the most efficient and robust type of RL algorithms are continuous state actor-critic algorithms that have the advantage of fast learning and the ability to generalize to new and unseen traffic conditions. These algorithms are utilized in this paper to design adaptive traffic signal controllers called actor-critic adaptive traffic signal controllers (A-CATs controllers).The contribution of the present work rests on the integration of three threads: (a) showing performance comparisons of both discrete and continuous A-CATs controllers in a traffic network with recurring congestion (24-h traffic demand) in the upper downtown core of Tehran city, (b) analyzing the effects of different traffic disruptions including opportunistic pedestrians crossing, parking lane, non-recurring congestion, and different levels of sensor noise on the performance of A-CATS controllers, and (c) comparing the performance of different function approximators (tile coding and radial basis function) on the learning of A-CATs controllers. To this end, first an agent-based traffic simulation of the study area is carried out. Then six different scenarios are conducted to find the best A-CATs controller that is robust enough against different traffic disruptions. We observe that the A-CATs controller based on radial basis function networks (RBF (5)) outperforms others. This controller is benchmarked against controllers of discrete state Q-learning, Bayesian Q-learning, fixed time and actuated controllers; and the results reveal that it consistently outperforms them.     

Identifying trip ends from raw GPS data with a hybrid spatio-temporal clustering algorithm and random forest model: a case study in Shanghai

ABSTRACT

Smartphones have been advocated as the preferred devices for travel behavior studies over conventional surveys. But the primary challenges are candidate stops extraction from GPS data and trip ends distinction from noise. This paper develops a Resident Travel Survey System (RTSS) for GPS data collection and travel diary verification, and then uses a two-step method to identify trip ends. In the first step, a density-based spatio-temporal clustering algorithm is proposed to extract candidate stops from trajectories. In the second step, a random forest model is applied to distinguish trip ends from mode transfer points. Results show that the clustering algorithm achieves a precision of 96.2%, a recall of 99.6%, mean absolute error of time within 3?min, and average offset distance within 30 meters. The comprehensive accuracy of trip ends identification is 99.2%. The two-step method performs well in trip ends identification and promotes the efficiency of travel survey systems.     

A Reliable Adaptive High-Definition Transmission System for Distance Learning

There is an increasing demand for supporting high-quality real-time audiovisual services for the next generation wired and wireless networks. However, due to variety of bandwidths of different networks, it is a great challenge for deployment. In this paper, a novel high-definition (HD) video transmission system was proposed which depends upon reliable compound multicast protocols and QoS control over the various kinds of networks. This system detects client's network condition and assigns it to a proper proxy. Each proxy is capable of detecting network parameters and adaptively tuning such transport parameters as bit rate, video resolution, frame rate and QoS mechanisms to this condition. It also provides FEC error recovery under consideration of characteristics of MPEG4 video codec. Our simulation demonstrates that different network clients such as ADSL, CERNET, and CERNET2 can receive more video reliability with less delay.     

基于深度学习的车辆目标检测算法综述

车辆目标检测是自动驾驶环境感知的重要组成部分。近年来随着深度学习在目标识别领域取得重大突破,基于深度学习的车辆目标检测算法逐渐成为该领域的研究热点。论文对当前主流的两阶段车辆目标检测算法和单阶段车辆目标检测算法进行简要介绍,分析了其中几种具有代表性的卷积神经网络算法的优缺点,最后总结目前车辆目标检测存在的问题以及未来的发展方向。     

基于改进U型神经网络的路面裂缝检测方法

针对传统的裂缝分割算法难以识别狭窄裂缝且分割边缘不精准,从而造成识别精度较低的问题,研究了基于改进U型神经网络（Unet）的路面裂缝检测方法。由于传统Unet特征提取网络是层次较浅的浅层神经网络,难以提取更复杂的裂缝特征信息,故本文以牛津大学视觉几何组网络（VGG16）作为传统Unet的特征提取网络,提高网络的裂缝特征提取能力;为抑制高低阶特征融合时产生的无用特征,本文在模型解码部分添加压缩与激励单元（SE block）,构建裂缝注意力单元,使得网络可以关注不同通道下的裂缝特征,建立了基于SE block和VGG16的改进Unet网络（SE-VUnet）。研究采用迁移学习的方法,将在ImageNet上预训练好的VGG16网络权重迁移到裂缝检测中。通过挑选Crack500数据集,并使用摄像头采集图片构建1 600张路面裂缝数据集,再次训练SE-VUnet模型,获得裂缝区域分割结果。以查准率（precision）与查全率（recall）的加权调和平均值F1和雅卡尔（Jaccard）相似系数作为量化评价指标。将SE-VUnet分别与Unet、SOLO v2、Mask R-CNN以及Deeplabv3+进行分割效果和实时性对比。研究结果表明：SE-VUnet模型的综合F1和雅卡尔系数分别为0.840 3和0.722 1,相比于Unet分别高出了1.04%和1.51%,且均高于其他3种对比模型;SE-VUnet的单帧图片预测时间为89 ms,在分割效果提升明显的情况下仅比Unet慢5 ms,优于其他模型。