Sequential visual place recognition using semantically-enhanced features

The problem of visual place recognition enables a vehicle to identify the place that it is currently at; that can be used to localize the vehicle through loop closures when the traditional SLAM algorithms fail due to low resolution of the sensors or extreme lightning conditions. We assume that the vehicle's route is filled with sequential places that the vehicle sees in the same order as it travels a pre-defined route. We handle three distinct challenges for this problem. First, since most places have a very small distinctiveness, we make the feature extraction algorithm conscious of the strong features like building facades and signboards. Second, there may be several dynamic entities that disallow reliable place recognition, and therefore we learn the model so as to eliminate the dynamic objects. Third, to minimize mis-classifications due to the non-distinctiveness of several places, we restrict the model to output smooth trajectories to minimize mis-classifications, attacking the sequential nature of the visual place recognition problem. The test domain may be somewhat different from the domain in which the place dataset is made, and we, therefore, use transfer learning to extract similar features. Experimental results show improved performance as compared to several state-of-the-art approaches.