Navigating among people in crowded environment

We present a novel set of data for the evaluation of visual place recognition in both indoors and outdoors environment in addition to sensor information to evaluate human-robot interactions in crowded areas. The datasets were recorded in the Royal Alcázar of Seville (Spain). We recorded a large set of images sequences from a stereo camera and scan measurements from three laser mounted on a moving robot.

The datasets are timestamped and stored by means of the well-known Robot Operating System (ROS) log functionality. The robot traveled more than one kilometer in each experiment, and every trial was performed at different time of the day so we could capture the evolution of lighting conditions over the images. The tourist attendance also depends on the hour, providing datasets with a lot of examples to model into a social-way the different places such as corridors, gates, queues, groups of people, etc.

A data paper on these datasets is available here

Each of the 10 datasets are stored separately in three files: one for sensors measurements, other for raw RGB images and a third one with the rectified grayscale images. All the datasets are logged and also processed using ROS tools, such as the ROS Bag. The different information stored into the logs also follows ROS interfaces and development main guidelines, so that the reader can understand easily the dataset with minimum ROS background.

The three log files per dataset contain the following information:

• • Robot Odometry


• • Laser Measurements


• • Tranformations between Sensors


• • Camera Info


• • Raw RGB Image


• • Rectified Image

The robot platform used for the datasets is a Pioneer 3AT with a simple aluminum structure to place the sensors and the computer:

• • A stereo camera facing forward at 1.2m height.
• • Two Hokuyo UTM-30LX placed parallel to the floor facing forward and backwards.
• • A Hokuyo URG-04LX tilted 30º in front of the robot.
• • Encoders in the robot base for odometry computation.

The datasets presented in this work represent a valuable set of images and range measurements that could be used to the comprehension of several human-robot behaviours or interactions.

The scenario considered in this work is a crowded mostly planar area. Tilted laser is aggregated to be used for small steps and ramps detection and determine by software if the robot will be able to cross the area or not. Images gathered with both left and right cameras make the dataset interesting for testing loop closure and kidnapping algorithms based in scene recognition. These datasets offer the possibility to study daylight variation in a mixed indoor outdoor structured environment and use this information for improving localization or augmenting accuracy. This dataset is also useful for testing algorithms for removing illumination effect in scenes.