LiDAR Place Recognition with Image Retrieval

This thesis is about LiDAR place recognition. Place recognition is the problem of being able to recognize already seen or visited places – an important sub-problem of robot navigation. LiDAR sensors offer accurate and cost-effective range and reflec-tivity data that can replace or complement RGB cameras.

Place recognition has been studied with different sensors and methods for many years. Traditional methods use handcrafted features to match images in order to recognize places. In recent years, the surge of deep learning has made learned features the main approach.

In this work LiDAR place recognition is studied with exported 2D pixel images and deep learning models. Place recognition is posed as an image retrieval problem, where a model is trained to learn a feature space in such a way that the similarity of images can be conveniently compared. With a trained model, one can use an image to search for other similar images, and thus recognize places.

The key finding of the thesis publications is that place recognition with image retrieval using exported pixel images from LiDAR scans is a well performing method, as evidenced by achieving about 80% recall@1 with 5 meter test distance in urban outdoors and 1 meter indoors. The other key findings are: Loop points in the route are detectable with image retrieval type methods. LiDAR is a competitive modality versus RGB. LiDAR depth maps are more robust to change than intensity maps. Generalized mean is a good pooling method for place recognition. Simulated data is beneficial when mixed in with real-world data at a suitable ratio. Dataset quality is very important in regards to ground truth position and LiDAR resolution.