3D Object Detection in Urban Environment

A 3D object detection system is developed on our vehicle-borne sensing platform POSS-v. Unlike most similar approaches that follow a bottom-up procedure (Segmentation, Labeling & Merging), our approach focus on a subset of sensing data labeled as "salient". The salient data belong to some pre-defined classes of objects, and are highlighted by a graph matching procedure.

The dataset used in this research is collect by a SICK LMS-200 LiDAR mounted on the right-side of the vehicle's roof, with the ground data removed already. It contains a variety of objects appearing frequently in urban scenes. Data files are provided in ascii format: x y z label, one point per line, space as separator. We manually labeled 16 classes of objects (the left parts are undefine), which are listed below. Download link is provided here.

Referred publications:

• Computing Object-based Saliency in Urban Scenes Using Laser Sensing,

Y. Zhao, M. He, H. Zhao, F. Davoine, H. Zha, IEEE Int. Conf. on Robotics and Automation (ICRA), 2012. (paper, presentation)

• Moving Object Detection and Tracking at Intersections using a Network of Laser Scanners,

H. Zhao, J. Sha, Y. Zhao, J. Xi, J. Cui, H. Zha, R. Shibasaki, IEEE Tran. on Intelligent Transportation Systems, 2012. (paper)

• Scene Understanding in a Large Dynamic Environment through a Laser-based Sensing,

H. Zhao, Y. Liu, X. Zhu, Y. Zhao, H. Zha, IEEE Int. Conf. on Robotics and Automation (ICRA), 2010

• Segmentation and Classification of Range Image from an Intelligent Vehicle in Urban Environment,

X. Zhu, H. Zhao, Y. Liu, Y. Zhao, H. Zha, IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), 2010

Trajectory Optimization of Multiple Object Tracking

For applications like trajectory analysis, scene modeling and abnormal detection, trajectories of good quality are required. However, for current moving object tracking systems, split/merge events happen frequently and reduce trajectory quality greatly. Therefore, a trajectory processing method that improves the quality of tracking results is necessary.

Referred publications:

• Moving Object Trajectory Processing Based on Multi-Laser Sensing,

Y. Zhao, H. Zhao, J. Sha, H. Zha, IEEE Int. Conf. on Intelligent Transportation Systems (ITSC), 2011. (paper, presentation, video)

• Trajectory Analysis of Moving Objects at Intersection Based on Laser-Data,

J. Sha, Y. Zhao, W. Xu, H. Zhao, J. Cui, H. Zha, IEEE Int. Conf. on Intelligent Transportation Systems (ITSC), 2011. (paper, presentation)

Other Works

Demands on real time lane detection & tracking are great for robotics, ADAS and autonomous vehicles. A lane detecion & tracking system is developed and tested on our vehicle platform POSS-v, also a cross-line alerting module is introduced. Some demo videos can be downloaded here(lane detection, lane tracking.avi).

We also develop a web-based interface for manualy labeling traffic lanes from video data, using the Bing Mapping SDK. A demo page can be visited, and some sample data can be found at 40.0364N,116.306E. Also some demos of POI extraction interface are provided(line label.avi, plane label.avi).