Life - Long Place Recognition by Shared Representative Appearance Learning Fei Han 1 , Xue Yang 1 , Yiming Deng 2 , Mark Rentschler 3 , Dejun Yang 1 , and Hao Zhang 1 1. Colorado School of Mines 2. University of Colorado Denver 3. University of Colorado Boulder Motivation Approach • We address the critical long-term place recognition task with strong appearance variations due to changes of illumination, vegetation, weather, etc. Hao Zhang, Ph.D. Assistant Professor Division of Computer Science Colorado School of Mines Phone: (303) 273-3581 Email: [email protected] HCRobotics Lab: http://hcr.mines.edu Contact Results • Our method achieves the state-of- the-art long-term place recognition performance, and outperforms baseline (feature concatenation) and previous (SeqSLAM, BRIEF- GIST, Color and LBP) methods . Scene Representation • A set of heterogenous visual features are utilized to capture image information and represent scenes. Summary • We propose an innovative long-term place recognition method based on Shared Representative Appearance Learning (SRAL) that is robust to strong appearance changes. Optimization Formulation • Place recognition tasks are formulated as a regularized sparse optimization • ℓ 2,1 -norm regularization enforces sparsity of each row of • ℓ M -norm regularization enforces sparsity between different feature modalities Soundness • Algorithm 1 converges to the global optimal solution Notation X: scene feature matrix; Y: scene indicator matrix; W: weight matrix Spring Summer Autumn Winter Morning Afternoon October December • Single feature modality or simple feature concatenation scheme cannot well represent the same place with appearance changes. Our method can learn and fuze multimodal features to build highly discriminative place representations that are robust to appearance changes. Spring Summer Autumn Winter Color GIST HOG LBP Raw Nordland Dataset (Different Seasons) Precision-recall curve Feature weight Precision-recall curve Feature weight St Lucia Dataset (Various Times of the Day) CMU-VL Dataset (Different Months) Place Recognition with Fusion • Feature weight • Feature fusion for image matching Precision-recall curve Feature weight where is the matching score