1. Introduction

        2. Demo

        3. Publication

        4. About us

 

The R2-D2 system has been renamed as LEADER.

R2-D2 now can run on Malacca Ship trajectories.

The demo paper of R2-D2 has been accepted by VLDB 2013!

The research paper of R2-D2 has been accepted by KDD 2013!

1. Introduction


Short Demo Video (50 s, only in Singapore Taxi Trajectories)

If you are from China, please click here :).

Path prediction is presently an important area of research with wide applications. However, most of the existing solutions are based on eager learning methods which commit to a model or a set of patterns extracted from historical trajectories for prediction. Problems can arise in dynamic environments where the underlying models change quickly or in regions which are not covered with statistically significant models or patterns.

In this demonstration, we present the “R2-D2” system for supporting probabilistic path prediction in dynamic environments. The core of our system is a “semi-lazy” approach to probabilistic path prediction, which builds prediction model on the fly using historical trajectories that are selected dynamically based on the trajectories of target objects. Our demo provides a visual interface to illustrate probabilistic path prediction on several real-world datasets. The demo also provides the ability to interactively set test scenarios with regard to the input datasets and various parameters, which gives users the opportunity to see the distinctive features of our system. Users will also be able to see how the prediction results of our system compares against other competitors through the statistics that we display.

Architecture

Figure 1: Architecture of the "R2-D2" System.

"R2-D2" has a " Update" process ( blue solid lines) and a " Predict" process ( red dotted lines). "Predict" process has two sub-process: "Lookup" process and "Construct" process. Figure 1 depicts the architecture of the "R2-D2" system, which has two main components: the Trajectory Grid (TG) and the Prediction Filter (PF). Corresponding to TG and PF, there are two separate processes: "Update" and "Predict". The "Update" process continuously collects streaming trajectories from the dynamic environment. TG buffers the streaming trajectories of moving objects (blue cars in Figure 1). The "Predict" process makes path prediction. This process can be divided into two sub-processes: "Lookup" and "Construct". In Figure 1, we want to predict the path of Op (the red car). In "Lookup", we use h-backward trajectory of Op to retrieve reference objects from TG. In the "Construct" process, the reference objects retrieved in "Lookup" are used to construct a model for making path prediction.

Interface

(a) Screenshot of the main interface

(b) Screenshot on Singapore Taxi data

(c)Screenshot on Human Tracking data

(d)Screenshot on Brinkhoff Oldenburg data

Figure 2: Main interface of the "R2-D2" System.

Fig. 2 is a screenshot of the demo interface. It consists of two parts: control pane (left part) and display pane (right part). The control pane is composed of four areas from top to bottom: (A) Data sets, (B) Prediction setting, (C) Output and (D) a group of control buttons. In the area (A), users can select the test data sets. In the area (B), the user can set various parameters; specifically, users can set the confidence threshold and enable the self-correcting continuous prediction, both of which are distinct features of our system. In the area (C), the users can see the different statistics information of our prediction method compared against competitors.

The display pane is composed of two parts from top to bottom: the canvas view and the timeline bar. In canvas view, the user can see the map of visualized trajectory, moving objects and predicted path. In Fig. 4, the the red points represent the target objects whose path need to be predicted, and the sequences of green dots are the predicted path. The background blue map is the visualization of trajectories of other objects. The timeline bar lets us set the time interval for selecting datasets.

back

2.Demo


Please see our demo here. (The demo is based on JDK 7u17. You have to update your JDK up to 7u17)
 

back

3. Publication



 

back

4. About us


School of Computing, National University of Singapore

Associate Professor: Anthony K. H. Tung

PhD Student: Jingbo Zhou

Data Analytics Department,Institute for Infocomm Research, A*STAR, Singapore

Dr. Wei Wu

Dr. Wee Siong Ng

Contact us

Please send email to: jzhou AT comp. nus .edu .sg


back