Evaluation of Image-Based Landmark Recognition Techniques

Recognizing landmarks in sequences of images is a challenging problem for a number of reasons. First of all, the appearance of any given landmark varies substantially from one observation to the next. In addition to variations due to different aspects, an illumination change, external clutter, and changing geometry of the imaging devices are other factors affecting the variability of the observed landmarks. Finally, it is typically difficult to make use of accurate 3D information in landmark recognition applications. For those reasons, it is not possible to use many of the object recognition techniques based on strong geometric models. The alternative is to use image-based techniques in which landmarks are represented by collections of images which capture the “typical” appearance of the object. The information most relevant to recognition is extracted from the collection of raw images and used as the model for recognition. This process is often referred to as “visual learning.” Models of landmarks are acquired from image sequences and later recognized for vehicle localization in urban environments. In the acquisition phase, a vehicle drives and collects images of an unknown area. The algorithm organizes these images into groups with similar image features. The feature distribution for each group describes a landmark. In the recognition phase, while navigating through the same general area, the vehicle collects new images. The algorithm classifies these images into one of the learned groups, thus recognizing a landmark. Unlike computationally intensive model-based approaches that build models from known objects observed in isolation, our image-based approach automatically learns the most salient landmarks in complex environments. It delivers a robust performance under a wide range of lighting and imaging angle variations.