Implementation and Evaluation of Image Sequence Based Place Recognition Utilizing a Humanoid Robot
In the domain of image based localization, alternatives to image feature based algorithms have been developed making use of similarity metrics that operate directly on pixel intensity values. These approaches can work robustly and efficiently in cases when the recognition process needs to deal for example with varying lighting conditions or changes in scenery detail. In addition to employing such a pixel based similarity metric a recent algorithm called OpenSeqSLAM processes image sequences instead of single images to improve the recognition. However OpenSeqSLAM is not robust in cases of perspective change. This makes it problematic to use in many robotic applications when the camera perspective is not fixed to certain positions with fixed orientations. In this contribution an approach is developed and evaluated that aims to mitigate the effect of perspective change on recognition performance by combining the concept of OpenSeqSLAM with an alternative similarity metric called tangent distance. It was further analyzed if the algorithm can be suitably designed to run on a humanoid robotic platform and how it can utilize the robots capabilities. To enable evaluation a test application called Dream Viewer for image sequence based localization algorithms has been developed. In result of the first tests an adapted and heuristic version of the algorithm was developed and evaluated as well. This algorithm aims to run on hardware performance constraint robotic embedded systems. Developed algorithms were tested offline with recorded image data as well as online on a humanoid robot platform called Myon. Results indicated that the developed algorithms using tangent distance can perform superior in terms of recognition performance compared to the standard OpenSeqSLAM algorithm in the tested cased of perspective change.