Vision-based global localization without a prior location estimate is a fundamental task for safe and efficient vehicle navigation in GPS-denied environments. Cross-season localization, in which query and database images involve different seasons is one of the most challenging task scenarios, owing to appearance variations among seasons. Because of recent advances in deep convolutional neural networks (DCNs) and transfer learning techniques, the task can be solved accurately by training and fine-tuning a DCN-based visual place classifier. However, the direct implementation of this would require collecting and storing a large amount of visual experiences (i.e., training data) for every new season, which is impractical. The goal of our study is to suppress the space cost for long-term memory and to develop a constant cost framework for long-term global localization. Moreover, we formulate and consider the task of experience compression as a scheduling problem of how to choose the part of the previous season's experience that is to be replaced with the current season's experience, to achieve an optimal tradeoff between localization accuracy and training efficiency. Experimental results using the publicly available North Campus Long-Term autonomy dataset validate the efficacy of our proposed approach.
Long-Term Vehicle Localization Using Compressed Visual Experiences
01.11.2018
1078888 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
LONG-TERM VISUAL TRAILER TRACKER FOR VEHICLE-TRAILER ANGLE ESTIMATION
Europäisches Patentamt | 2022
|LONG-TERM VISUAL TRAILER TRACKER FOR VEHICLE-TRAILER ANGLE ESTIMATION
Europäisches Patentamt | 2024
|