Visual place recognition is a core capability needed by modern mobile robots to perform long-term Simultaneous Localization And Mapping (SLAM). The goal of place recognition is to localize a mobile robot in an environment by identifying previously visited locations. Long-term place recognition is a challenging problem due to perceptual aliasing and long-term appearance changes (i.e. seasonal changes), along with traditional challenges including occlusions, dynamic objects (e.g., cars and pedestrians), and illumination variations. The objective to collect this new dataset, named Multimodal Omni-directional Long-term Place-recognition (MOLP), is to investigate and evaluate long-term place recognition based on multisensory omni-directional observations.
The MOLP dataset is collected using an omni-directional stereo camera that is installed on a car, as shown in Figure 1. A GPS module is used to collect ground truth of locations for validation of place recognition results. The dataset comprises of an omni-directional intensity image, an omni-direction depth image, and GPS data in each frame. Data was collected from two routes (Route A and Route B) in the areas of Golden, Colorado. Route A is a loop around the city of Golden covering the downtown and Colorado School of Mines campus. Route B is mountainous route up the Clear Creek Canyon road. The car was driven in clockwise direction or counterclockwise direction during two different times of the day (morning and evening), different months, and different seasons. The example video in Figure 2 shows the counterclockwise direction of Route A in the morning, and the video in Figure 3 shows the counterclockwise direction of Route B in the evening.
Downloads and Citation
The full MOLP dataset will come soon. The dataset provided here is for non-commercial research/educational use only.
Questions and Suggestions
Please contact Ashwin Mathur: mathur -AT- mymail -DOT- mines -DOT- edu