Bayesian place recognition based on bag of objects for intelligent vehicle localisation. Issue 11 (15th August 2019)

• Record Type: Journal Article
• Authors: Wang, Xianglong; Hu, Zhaozheng; Tao, Qianwen; Huang, Gang; Mu, Mengchao
• Is Part Of: IET intelligent transport systems. Volume 13:Issue 11(2019)
• Page Start: 1736
• Page End: 1744
• Publication Date: 2019-08-15
• Subjects: probability -- image recognition -- image sequences -- object detection -- feature extraction -- image matching; place recognition efficiency -- classic BoW -- intelligent vehicle localisation -- pre‐built image sequence -- visual map‐based vehicle localisation -- Bayesian place recognition algorithm -- classic bag -- object probabilities -- pre‐trained AlexNet -- BoO feature -- scene image similarity -- appearance feature -- motion clue -- dominant objects -- multiscale localisation -- bag of objects -- classic bag of words -- dynamic illumination environments -- KITTI datasets; Intelligent transportation systems -- Periodicals; Electronics in transportation -- Periodicals; 388.31205
• DOI: 10.1049/iet-its.2018.5431
• Languages: English
• ISSNs: 1751-956X

Abstract

Abstract : Place recognition matches a corresponding image from a pre‐built image sequence, and it plays an important role in visual map‐based vehicle localisation. This study proposes a Bayesian place recognition (BPR) algorithm based on a bag of objects (BoO). The proposed BoO is superior to the classic bag of words (BoW), with a faster speed and greater robustness for scene representation. Like BoW, BoO uses object probabilities computed from the pre‐trained AlexNet to construct a histogram‐like vector called the BoO feature. A small portion of objects, named dominant objects, are computed from map data to reduce the storage requirement. The BoO feature is used to describe the scene image similarity. Owing to the excellent recognition performance of AlexNet, the BoO feature is robust to dynamic illumination environments. With BoO, the BPR model is employed to fuse appearance feature and motion clue for place recognition. Moreover, a multi‐scale localisation strategy is developed to improve the place recognition efficiency. The proposed method was tested with the collected data and the KITTI datasets. The method achieved 0.37 frame errors, taking an average of 4.7 ms for place recognition. Experimental results demonstrated that the proposed method outperforms the classic BoW and binary BoW.