This research is to integrate communication, traffic and driver behaviours, vehicle dynamics, and environment conditions in a unison framework to increase safety, reliability, and performance of connected and automated vehicles within a mix traffic condition.
To achieve this goal, TUPA has been developing an architecture that leverages existing tools like VISSIM, Driving simulator, Virtual Reality and Python to analyze and simulate multiple aspects of traffic including vehicles and pedestrians.  This tool will enable to develop and validate techniques to improve traffic congestion, safety, and network performance.

In designing scenarios for the integration of Vissim and Driving simulator, a robust algorithm has been successfully implemented in the platform. This algorithm adapts its behaviour (traffic flow in transport) autonomously in response to the variation of network conditions.

Having human in the loop setting, alternative route guidance information will be displayed in driving simulator so that driver’s decision if he or she is willing to change the route based on the information will be recorded. In this process, the subject vehicle in the simulator and surrounding traffic in VISSIM will interact each other. In addition, all recoded outcomes from the driving simulator will be input to traffic simulation for the large network evaluation. The developed traffic simulation will identify typical traffic indicators such as delay and extreme delay, queue and the number of stops based on the scenarios developed in the driving simulator.

This research keeps developing with the KAIST team since middle of 2019. Transportation experts are aware that it is urgent to take measures to cope with mixed traffic between autonomous vehicles and conventional vehicles, which will be indispensable in the future. It is expected that this new traffic flow not only directly and indirectly leads to traffic accidents but also has a serious side effect on traffic efficiency. Until a fully autonomous vehicle occupies the road, new traffic control techniques are needed. In this study, we propose a new traffic signal controller that supports autonomous driving by driving the driver safely through the virtual environment. This study is expected to be synergistic effect of performance evaluation for future traffic in conjunction with the Connected ITS project and the K-City autonomous vehicle test bed currently underway in Korea. It is expected that many universities and research institutes will benefit from this research since Korea has not developed a human participatory machine learning platform using virtual reality or is in the very early stage.

The dashboard below was developed through Elastic open source software using the Seoul metro passenger flow data in 2014.

Data visualization has been important in democratizing data and analytics and making data-driven insights available to workers throughout an organization. Data visualization also plays an important role in big data and advanced analytics projects. As a transportation field accumulated massive troves of data during the early years of the big data trend, they needed a way to quickly and easily get an overview of their data. Visualization tools were a natural fit.

Visualization is central to advanced analytics for similar reasons. When advanced predictive analytics or machine learning algorithms are available, it becomes important to visualize the outputs to monitor results and ensure that models are performing as intended. This is because visualizations of complex algorithms are generally easier to interpret than numerical outputs.
In TUPA, we utilized one of the strongest searching engines called Elasticsearch to make this visualization works. The flowchart below shows the process to deal with big data and visualize it.

For more detailed information please contact our TUPA members below;
Xu Yanping, 1394593741@qq.com
Cheng Lyu, lvcheng1996@126.com

We utilized TurtleBot3 which adopts ROBOTIS smart actuator Dynamixel for driving.

TurtleBot3 is a ROS-based mobile robot. We customized it to reconstruct the mechanical parts and use optional parts such as the computer and sensor. SLAM, Navigation and Manipulation, makes it to build a map and can drive around the room. Also, it can be controlled remotely from a laptop, joypad or Android-based smart phone.

The project allows the robot to detect the lane(s) and obstacles to avoid. Various algorithms such as SLAM, CNN, LSTM and OpenManipulator are embedded in the robot to make better robot behavior.

The following video demonstrates the navigation function.

For more detailed information please contact our TUPA members below;
Taeho Oh, taeho.oh.0219@gmail.com
Xu Bicheng, 747294301@qq.com

Research Aim

To provide empirical evidence on the relationship between built environment and public sharing bike flow in Suzhou, China.

Research Objectives

• To examine the global impacts of built environment on public sharing bike flow.
• To understand the effects of spatial variation of those built environment on public sharing bike flow

Study Area

The study area of this research focuses on Suzhou located in the southeast Jiangsu Province of East China and east about 100 km to Shanghai (Figure 1(A)).

There are around 1,750 bike stations and 40,000 public sharing bikes put into use in Suzhou (Figure 1(B)).

Fig. 1: Study area. (A) Location of Suzhou in China; and (B) the spatial distribution of bike stations, metro stations and population density in urban area of Suzhou.

Data

Data Sources

Operationalization of Variables

Methodology

Global Regression

Geographically Weighted Regression (GWR)

GWR is a local regression model. Coefficients are allowed to vary.

Bi-squared Weighting Function

Results and discussions

Global Regression

Table 1: The results of Global Regression.

Note: Values in bold are significant at 0.1 level.

Geographically Weighted Regression (GWR)

Fig. 3: Comparisons of explanatory power of Global regression and GWR.

Fig. 4: Spatial distributions of local coefficients on working day and t-value with significance less than 90%.

Conclusions

The goodness of fit in the GWR is better than the global regression model.

Acknowledgements

This work was supported by Jiangsu Industrial Technology Research Institute and Research Institute of Future Cities at Xi’an Jiaotong-Liverpool University.

For more detailed information please contact our TUPA members below;
Chunliang Wu, Chunliang.Wu@monash.edu

INTRODUCTION

Transportation data is of great importance for intelligent transportation system. Missing data problems are inevitable during data collection.

Challenges in existing imputation methods: potential useful information is not efficiently used in the modeling process; methods considering temporal correlation usually assuming that linear relationships exist between observed variables and latent variables; most techniques fail to measure the uncertainty.

This study introduces the use of a self-measuring multi-task Gaussian process (SM-MTGP) method for imputing missing data.

CONTRIBUTIONS

A SM-MTGP method is proposed to combine features from tasks and inputs to measure similarities jointly.

Dependencies of tasks and inputs are explored via covariance functions under SM-MTGP framework.

Correlations between responses are captured to provide additional information for enhancing imputation accuracy.

METHODOLOGY

Brief review of MTGP

Assuming we have \(Q\) tasks and a set of observations \(Y = \left\{ {{y_{i1}},{y_{i2}}, \ldots {y_{iD}}} \right\}, i = 1, \ldots ,Q\), for each corresponding task at \(????\) distinct inputs, where \(????_{????????}\) is the response for \(????^{????ℎ}\) task given the input \(????_????\).

FIGURE 1 Vectorization of matrix Y

When the SM-MTGP model is introduced to the imputation of missing values of transfer passenger flow, the shared information of tasks is considered in terms of the temporal relatedness of various days. Transfer passenger flow over \(Q\) days can be treated as \(Q\) tasks, and the number of sampling time intervals \(D\) per day represents \(D\) distinct inputs. We define a matrix \(Y = \left\{ {{y_{ij}}} \right\}(i = 1,2,…,Q;j = 1,2,…D) \in Q \times D \), where \({{y_{ij}}}\) is number of transfer passengers for the \({i^{th}}\) day (task) on the \({j^{th}}\) time interval (input). By stacking the column vectors of \(Y \in Q \times D \), a \(Q \times D\) dimension vector \({\bf{y}} = vec(Y)\) is obtained (Figure 1).

The MTGP model of \({{\bf{\tilde y}}}\)can be described as Equation (1):
$${{\tilde y}_{ij}} = {m_{ij}} + \varepsilon ,\quad \varepsilon \sim N\left( {0,{\sigma ^2}} \right) \tag{1}$$ where \({m_{ij}}\) is the expected value of the element \({{\tilde y}_{ij}}\), and \(\varepsilon\) is an additive Gaussian noise with variance \({{\sigma ^2}}\).
$$m \sim N\left( {0,{\Sigma _Q} \otimes {\Sigma _D}} \right) \label{TGP} \tag{2}$$ $${\Sigma _Q} = K_Q^fG_Q^m, \quad{\Sigma _D} = K_D^fG_D^m \label{covariance matrix} \tag{3}$$ The covariance matrices \({\Sigma _Q}\) are defined as a product of kernel of days features (tasks) \(K_Q^f\) and the self-measuring kernel \(G_Q^m\), and \({\Sigma _D}\) are defined as a product of the kernel of time intervals features (inputs) \(K_D^f\) and self-measuring kernel \(G_D^m\).
$${K_Q^f} = k\left( {{y_i},{y_j}} \right) \in \mathbb{R}^{Q \times Q}, \quad {G_Q^m} = g\left( {{y_{i:}},{y_{j:}}} \right) \in \mathbb{R}^{Q \times Q} \tag{4}$$ $${K_D^f} = k\left( {{y_h},{y_l}} \right) \in \mathbb{R}^{D \times D}, \quad {G_D^m} = g\left( {{y_{:h}},{y_{:l}}} \right) \in \mathbb{R}^{D \times D} \tag{5}$$ where \(k\left( {{y_i},{y_j}} \right)\) and \(k\left( {{y_h},{y_l}} \right)\) indicate covariances of features of \({i^{th}}\) day and \({j^{th}}\) day, and covariances of features of \({h^{th}}\) time interval and \({l^{th}}\) time interval, respectively. Similarly, \(g\left( {{y_{i:}},{y_{j:}}} \right)\) and \(g\left( {{y_{:h}},{y_{:l}}} \right)\) measure covariances of self-measuring observations of \({i^{th}}\) day and \({j^{th}}\) day, and covariances of self-measuring observations of \({h^{th}}\) time interval and \({l^{th}}\) time interval.

By following the principle of MTGP, the joint distribution of \({\tilde Y}\) can be described as Equation (6), where \(\Phi = {\Sigma _Q} \otimes {\Sigma _D} + {\sigma ^2}{\bf{I}}\).
$$\int {p\left( {\tilde Y|M,0,{\sigma ^2}} \right)} p\left( {M|{\Sigma _Q},{\Sigma _D}} \right)dM = N\left( {{\bf{\tilde y}}|{\bf{0}},\Phi } \right) \tag{6}$$ Using a Gussian process framework given the observed number of transfer passengers, the unobserved passenger flows in \(Y\) can be derived by predictive equation (7). $$E[{{\tilde y}_{ab}}|{{{\bf{\tilde y}}}_{obs}},{\Sigma _Q},{\Sigma _D}] = \left( {{\Sigma _{{Q_a}}} \otimes {K_{{D_b}}}} \right)_{obs}^T\Phi _{obs}^{ – 1}{{{\bf{\tilde y}}}_{obs}} \tag{7}$$ where \({\Phi _{obs}} = {\bf{P}}\Phi {{\bf{P}}^T} \in \mathbb{R}^{M \times M} \) is a covariance matrix over the observed transfer passenger flows in \(Y\), \({{\Sigma _{{Q_a}}}}\) denotes \({a^{th}}\) column vector in \({\Sigma _Q}\), which measures the similarities between \({a^{th}}\) day and all the other days among \(Q\) days, and \({{K_{{D_b}}}}\) indicates \({b^{th}}\) column vector of \({K_D}\), which represents covariance between \({b^{th}}\) time interval and all the remaining time intervals of \(D\) samples.

EXPERIMENT

Data analysed includes 6-months of passenger flow data collected by WiFi sensors at Richmond railway station (Figure 2), Melbourne, Australia.

FIGURE 2 Map of location and train lines of Richmond station.

FIGURE 3 Map of 12 WiFi sensors distribution.

The deployed 12 sensors are distributed at platforms 7-10 and two sided underpasses (Figure 3).

IMPUTATION PERFORMANCE

Figure 4 indicates the RMSE results of discrete missing pattern with various algorithms. The improvements in RMSE by SM-MTGP is around 60% for three different missing rates.

FIGURE 4 RMSE Results for Different Missing Ratios of Discrete missing data.

Three mixed missing patterns under different missing ratios are reported (Figure 5-7). The SM-MTGP method is still able to obtain better performance compared with all the other methods, leading to improvements in RMSE up to 60%.

FIGURE 5 RMSE Results for Different Missing Ratios of Mixed Missing Data
with One Random Day Missing.

FIGURE 6 RMSE Results for Different Missing Ratios of Mixed Missing Data
with Two Random Day Missing.

FIGURE 7 RMSE Results for Different Missing Ratios of Mixed Missing Data
with Four Random Day Missing.

CONCLUSIONS

Imputation accuracy can achieve around 60% improvement in RMSE in all the tested missing scenarios compared with the base model.

SM-MTGP significantly outperforms other methods under the large missing ratio.

On-going research on incorporating other features into this algorithm to make application on large-scale transit network and simplifying model computational complexity.

For more detailed information please contact our TUPA members below;
Wenhua Jiang, wenhua.jiang@monash.edu

Intro

It’s not easy to pick up this background image for the post. It shows the Manhattan Peninsula, New York, and the sky reflects the buildings on the ground. This fantasy scene reminds me the complex relationship between space and time, which closely related to the topic of this article: Spatiotemporal forecasting of traffic by using 3d convolutional neural networks.

On December 31, 2014, a deadly stampede occurred in Shanghai, near Chen Yi Square on the Bund, where around 300,000 people had gathered for the new year celebration. 36 people were killed and there were 49 injured, 13 seriously (Wikipedia). From the follow-up reports, it can be known that Tencent’s user online information has roughly detected that the traffic in the area was too dense. So data from social media and cell phone signal can infer the regional crowd density. If the corresponding predictions and analysis can be made, such tragedies will not happen.

Traffic forecasting has been studied for decades. There are many outcomes of models and theories. For the regional prediction, it’s a prevalent way to split the research area into grids and analysis them by computer vision models. Each square in the girds just like the pixel in an image.

Zhang (2016) presented the classic Deep-ST model. It treats the research area as image and the predicted result combined with the convolutional models from different periods. However, the convert not only limited to traffic volume. The vehicle speed can also be transformed into images. The images below shows the traffic speed representation in a small-scale transportation network (Yu, et al 2017).

With the development of computer vision, deeper neural networks like ResNet has been presented recently. Zhang (2017) also upgraded the DeepST to ST-ResNet with ResNet models. However, the convolutional kernel in these models only focuses on spatial relations, not for a spatiotemporal space. It’s been proved that 3D Convolutional Networks can learn the spatiotemporal features (Tran 2015). But these improvements only appears in video related studies like behavior detection, human action detection, etc. So in this post, we will see the application of 3D ConvNets on traffic problems.

Convolutional operations

There are some different operations with the convolutional kernel in hidden layers. The following diagrams show the 2D convolutional operation with padding and dilation (Dumoulin and Visin, 2016 ). It’s different in 3D, but the ideas are same.

With the 3D convolutional operation, the kernel shape is 3 dimensional and it moves in 3 directions. Just like the animation below. For the transportation problems, the directions are latitude, longitude and time. In the model part, we also used padding and dilation with 3D kernels.

Data and Model

We take the bike sharing data in New York (BikeNYC) from the DeepST paper as an example here. Every circle stands for a station and the color means the number of bikes in dock. The research area is split into 16*8 grids, each square has in and out flow at a particular moment. The in and out flow are numbers of return and borrow bikes in the corresponding region.

The input is a time sequence from the time level of closeness, period and trend. They stand for the different extract frequency from raw data. If stack them together, the input shape would be X168. The X is the number of timesteps.

The model has three 3D convolutional layers and the flatten layer combines the external data like weather and holidays.

Experiments

I used Pytorch this time. The windows version just came out last month. It provides a lot of API and very easy to build a custom model structure. The author of ST-ResNet has opened his code, so we can reuse the dataset from Github.

To get the best kernel size, we take different combinations for training test. It turns out that 3*3*3 is also the optimal option for transportation data just like other papers have pointed out in some behavior detection tasks.

Although the model is much simpler than Deep-ST or ST-ResNet, it still achieved the best performance on BikeNYC and TaxiBJ datasets.

Visualization

Firstly, I have tried the Matplotlib in Python. The outcomes are good but they are static and not appropriate for the webpage. So I chose d3.js to draw the diagram from scratch. (The style with CSS and SVG tags is incompatible with this blog responsive CSS sheet. So I just give up displaying it on this page. Here is the gif version, simple and straightforward.)

Play around with this diagram: http://resuly.me/projects/3dconvs/

References

Zhang, J., Zheng, Y., Qi, D., Li, R., & Yi, X. (2016, October). DNN-based prediction model for spatio-temporal data. In Proceedings of the 24th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems (p. 92). ACM.

Zhang, J., Zheng, Y., & Qi, D. (2017, February). Deep Spatio-Temporal Residual Networks for Citywide Crowd Flows Prediction. In AAAI (pp. 1655-1661).

Yu, H., Wu, Z., Wang, S., Wang, Y., & Ma, X. (2017). Spatiotemporal recurrent convolutional networks for traffic prediction in transportation networks. Sensors, 17(7), 1501.

Tran, D., Bourdev, L., Fergus, R., Torresani, L., & Paluri, M. (2015, December). Learning spatiotemporal features with 3d convolutional networks. In Computer Vision (ICCV), 2015 IEEE International Conference on (pp. 4489-4497). IEEE.

Vincent Dumoulin, Francesco Visin – A guide to convolution arithmetic for deep learning

Part One —— 公共自行车在中国/苏州

公共自行车(BSs) 被公认为一种绿色、健康可持续的交通方式，并可提供与公交的最后一公里接驳服务。近年来，全世界范围的公共自行车发展迅速，截止2016年，已有超过1200个公共自行车系统部署在全世界范围，其中最大和发展最快的公共自行车国家就在中国——2016年底，中国已在430个城市和地区构建了公共自行车系统，意大利和美国紧随其后（无桩的共享单车不算在内）。 Read more

The lane change behaviour is not only impacted by the blockage but also the surrounding traffic.  The combination of these two impacts determines the subject vehicle lane change execution behaviour: i.e. to either Continue lane change execution or Not Continue lane change execution.

We developed the lane change execution model based on the model framework, followed by the lane change execution model (LCEM) calibration.  The lane change execution behaviour of PC is different from that of HV, it is worth exploring the LCEM for PC and HV separately. Based on the results of LCEM and analysis of the lane change execution characteristics of PC and HV, the LCEM for PC and HV are established and parameters are estimated accordingly.

Lane Change Execution Model Evaluation with Micro-Simulation Outcomes

Microscopic traffic simulation is an efficient and widely-used tool for analyzing the performance  of roadways on traffic, safety and the transportation system (Young, Sobhani et al. 2014).  The core of a simulation model is a set of mathematical algorithms that evaluate the motion of each vehicle on a second-by-second basis as it interacts with the road network, the traffic control system and the surrounding traffic environment.  A lane-changing model is incorporated as one of the basic driver behaviour interaction in the microscopic traffic simulations based on different algorithms.  Two major simulation models (VISSIM (2012)and AIMSUN (2012)) outputs of lane change processes are analyzed in this section.

• VISSIM

To assure a model reproduces real-world traffic conditions reasonably well, the simulation parameters have been calibrated using the data collected by the same camera. Headway is used as an indicator to adjust the parameter values.  When the Mann Whitney U test shows the distributions of headway from simulation and observation are comparable, the calibration process is completed. It is noted that after the calibration completed, VISSIM still provides the fixed value for lane change duration due to lack of lane change execution models integrated.

The output from Vissim shows that the duration of lane change is a fixed value.  For the passenger cars (PC), the lane change duration is 2.2s; and for the heavy vehicles (HV), the duration is 2.4s.  However, from the observed data collected in this study the lane change durations vary from 1s to 6.8s for both vehicle types.  The distributions of lane change duration from the observed data for PC and HV are shown.

Distributions of Lane Change Duration of PC and HV from Observation

 In terms of the lane change trajectory analysis, Vissim provides the lane change start and end locations which indicate the lane change trajectory in the longitudinal direction.  For passenger cars, the lane change length from Vissim ranges from 1.6m to 32.4m, while the observation of the lane change length ranges from 19.6m to 76.3m.  Figure 8-2 shows the distributions of the lane change length in longitudinal direction for passenger cars between Vissim output and observation. For heavy vehicles, the lane change length from Vissim ranges from 4.1m to 28.7m, while the observation of the lane change length ranges from 40.3m to 96.7m. Figure 8-3 shows the distributions of the lane change length in longitudinal direction for heavy vehicles between Vissim output and observation. The result implies that the lane change trajectory Vissim generates is much different from the real-life lane change trajectory for both passenger cars and heavy vehicles.

Distributions of Lane Change Length in Longitudinal Direction of PC

Distributions of Lane Change Length in Longitudinal Direction of HV

• AIMSUN

Figure shows the distributions of travel time during the blocked section for the passenger cars.  The travel time of passenger car from AIMSUN ranges from 2.6s to 16.8s, while the travel time of passenger car from observation ranges from 7s to 21.2s. It can be seen that the travel time is generally smaller than the observed travel time. Especially the minimal travel time from AIMSUN output is 2.6s which is 4.4s smaller than the minimal travel time of observations. It is not realistic for a PC to complete 140m in 2.6s.

Distributions of Travel Time of PC during the Blocked Section

Figure shows the distributions of travel time during the blocked section for heavy vehicles. The travel time of heavy vehicle extracted from AIMSUN ranges from 13.4s to 16s, while the travel time of passenger car from observation ranges from 9s to 21.6s.  It can be seen the travel time from observations for HV has a larger range while the travel time from AIMSUN output is clustered in a certain range, which means it does not reflect the observation well.

Distributions of Travel Time of HV during the Blocked Section

The results imply that the AIMSUN output differs from the observed data on lane change execution collected for this study, especially for the heavy vehicles.

Evaluation of Lane Change Execution Model

The lane change execution model is applied on the existing simulation outputs. This study assesses the results from lane change execution model and the real-life data to see the improvement of the model to the simulation output.

To apply the LCEM on the simulation output, the traffic output data need to be extracted from Vissim, identifying the subject vehicles and direct surrounding vehicles and recording the each vehicle traffic information in every 0.1s. A Matlab program is developed to apply the LCEM using the basic vehicle information from Vissim to predict the lane change behaviour. The result of lane change execution prediction (Continue Lane Change or Not Continue Lane Change) is thus used to identify the lane change execution durations.

The original output from Vissim shows that the duration of lane change is a fixed value, which is 2.2s for passenger cars.  The original simulation output is apparently not fitting the distribution of the lane change duration observation.  While applying the lane change execution model on Vissim, the lane change duration distribution is fitting the duration distribution of observations much better.

Figure shows the comparison of passenger car lane change duration distribution between the outputs applied the execution model, the real-life observed data and the outputs from simulations. The lane change duration of PC in Vissim is 2.2s and 0s in Aimsun. The comparison indicates that the lane change execution model is able to improve the simulation outputs to fit the real life lane change behaviour better for the passenger cars.

Comparison of Lane Change Duration Outputs for PC

The limitation of this work is all durations of the outputs are larger than 2.2s which is the fixed value provided from Vissim.  It can be seen that the range of 2~3s has very high portion in the model output compared to other duration ranges. The reason could be that it also includes the lane change durations which are smaller than 2.2s. This could be a good part of the future work to improve the current simulations.

Figure shows the comparison of heavy vehicle lane change duration distribution between the outputs applied the execution model, the real-life observed data and the outputs from simulations.  The lane change duration of HV in Vissim is 2.4s and 0s in Aimsun. The distribution of HV lane change duration is not as good as the distribution of PC lane change duration which fits the observed data better. However the trend of lane change duration is close to the observations than the outputs from the current traffic simulations. The result also indicates that employing the lane change execution model for HV would obtain the outputs closer to the real life lane change behaviour.

Comparison of Lane Change Duration Outputs for HV

Similar to the limitation for PC, the high portion of 2.5~3s of HV lane change duration could be caused by the fixed value of 2.4s which are extracted from Vissim.  Moreover, the size of observed data samples also could be a limitation of this work. The further work will need to collect more data for HV in different location.

Part One: Bike-share in China and Suzhou

Bike-share systems (BSs) are generally believed to be a sustainable and healthy transport mode and have the potential to enhance public transport by offering a last-mile service. The merging of BSs worldwide has been a relatively new trend. By the end of 2016, 1,286 BSs had been deployed worldwide, among which, the largest and fastest increasing BSs market is in China -by the end of 2016, China had established 430 BSs with a bike fleet of 1.9 million, making it the largest BSs market in the world, followed by Italy and USA, as shown in the following figure (note that the uprising dockless bike-share systems like Mobike are not included).

Top 10 bike-share system countries (by the end of 2016)

As for Suzhou Public Bike System (SPBS), it is a third generation bike-share system located in Suzhou – a major city (2,743 km2, 6.7 million population) located in East China. SPBS is operated by a private company Youon, under the supervision of Suzhou Urban Management Bureau (SUMB) – a local bike-share administration authority. It was first launched on August 30, 2010, with only 11 stations and 200 bikes. Currently, over 2,200 SPBS stations and 41,000 fleets are distributed all over the city. The average daily ridership of SPBS is more than 0.2 million in 2016, namely each SPBS bike used 4.7 times/day. SPBS contributes 0.32% of the overall transport modes.

Distribution of SPBS stations in Suzhou

Notes:
1. drag and scroll a mouse wheel to zoom in and out to search different areas
2. color:
Red circle – areas with stations >100
Yellow circle – areas with stations between 10 and 100
Green circle – areas with stations < 10

Access to SPBS was initially available to people who registered for a membership card. In 2016, Youon upgraded the system to allow smart phone access, so users can register, check in and out and receive information through an APP without a card. To some extent, it makes SPBS a third-half generation BSs.

For SPBS charging policy, the deposit is USD 30 for card registrations and free for APP registrations. The first hour’s ride is always free followed by USD 0.15 /h. No extra subscription fees or membership fees are needed. The financial requirements to gain access to the system are easy since cycling users are encouraged and subsided by government. Compared to western bike-share systems, great advantages could be seen in SPBS. See the following table.

Part Two: what do trip data tell us about SPBS

By using dataset provided by SUMB, we are able to get access to over 30 million SPBS trip records, which covers all 2,256 stations in Suzhou from June 1, 2016 to May 31, 2017. By analyzing this big data, tempo-spatial distribution, users’ characteristics and behaviors could be revealed.

Usage of SPBS in a random working day

Firstly, by dividing the picking up and returning activities, we could see how SPBS usage changes in a random working day in 24 hours in a city-scale. Two peak hours from 7 to 8 am and from 5 to 6 pm could be observed with unbalanced spatial distribution—— more pick-up activities could be observed in certain areas while other areas with more returning activities in peak hours.

Notes:
1. Every dote indicates a SPBS station
2. Different color means different numbers
3. Time can be seen in top left

SPBS trips dynamic assignment in a given weekday in a whole day

Since cycling in all SPBS stations can be tracked, a stations matrix is built up so that all the cycling trips could be assigned to existing road network. This following video shows how SPBS trips are assigned to road network in a random working day. A complete road network with all the public bike stations is built up and the assignment process is optimized so that cycling trips only exist in urban streets rather than expressways and freeways. Densest trips could be seen in roads in SIP, SND.

[jwp-video n=”1″]

Notes:
1. line width indicates trips numbers, thicker lines mean more bike trips
2. Time could be seen in bottom right

Average travel time by age and gender

Profiling method is used to analyze users characterizes. In the following two figures, trips from Jun 2016 to Jan 2017 are aggregated and trip time is distributed monthly in terms of different groups (by age and gender). It turns out that
young group (18-34 years old) and male group favor public bike more than other groups, while 5 min is the most popular trip time.

Notes:
1. legend is on the right, four age groups are divided
2. move the mouse on the figure and details would show out
3. Y-axis is ratio and X-axis is trip time.

Weekly trip frequency distribution by age

The following figure displays the weekly frequency of SPBS users in different age groups. Most people travel less than two times in one week – and younger the user are, more often they use public bikes.

Notes:
1. Legend is on the right, four age groups are divided
2. Move the mouse on the figure and details would show out

Year-trip distribution by age

This visualization shows that the total number of trip times of each month in different age groups from June 2016 to May 2017. People aged between 18 and 54 are the major user group of public shared bikes in Suzhou. In addition, the usage frequency of SPBS varies in different quarters. It indicates that the temperature and other environmental factors have a significant impact the usage of shared bikes. – for instance, in Jan and Feb (normally Spring Festival holidays), SPBS usage was greatly fluctuated and influenced.

Tips:
1. Legend is on the right, four age groups are divided
2. Move the mouse in each month line to see the detailed information

Access methods and users’ preference in a year

From 2016, users are able to get access to public bike system by scanning QR code through smart phone, which offers a convenient way to use public bike system. In this visualization, total trips grouped by two access methods (traditional smart card and QR) are given.
Big jump of QR users clearly indicates users’ preference on convenience.

Notes:
1. Legend is on the top right
2. Drag the bar at the bottom left and right to move horizontally to look close a chosen span

Access methods and users’ preference in a week

Two types of users (QR users and smart card users) are indicated and compared in a random working day in this figure. Interestingly, more people use smart card in the morning peak hour while scanning QR is more popular in night peak hour.