An Effective Subsuperpixel-Based Approach for Background Subtraction

Abstract

How to achieve competitive accuracy and less computation time simultaneously for background estimation is still an intractable task. In this paper, an effective background subtraction approach for video sequences is proposed based on a sub-superpixel model. In our algorithm, the superpixels of the first frame are constructed using a simple linear iterative clustering method. After transforming the frame from a colour format to gray level, the initial superpixels are divided into K smaller units, i.e. sub-superpixels, via the k -means clustering algorithm. The background model is then initialized by representing each sub-superpixel as a multidimensional feature vector. For the subsequent frames, moving objects are detected by the sub-superpixel representation and a weighting measure. In order to deal with ghost artifacts, a background model updating strategy is devised, based on the number of pixels represented by each cluster center. As each superpixel is refined via the sub-superpixel representation, the proposed method is more efficient and achieves a competitive accuracy for background subtraction. Experimental results demonstrate the effectiveness of the proposed method.

Fast-D: When Non-Smoothing Color Feature Meets Moving Object Detection in Real-Time

“…The local binary similarity pattern (LBSP) was a combination of intra-LBSP and inter LBSP used in [20], [21], which increased accuracy as well as complexity. The segmentation accuracy and complexity of the methods proposed by Hofmann et al [19] and and St. Charles et al [20], [21] was high compared with methods [9]- [17]. However, we see that none paid attention to find the trade-off between accuracy and complexity of the segmentation.…”

“…MOG [11] GMM [12] KNN [13] SBS [17] SuperBE [18] MOG2 [26] Cuevas [27] Berjón [28] Haines [29] SC-SOBS [30] WNN [31] CNN [32] DeepBS [33] RB-SOM [34] MSFgNet [35] MS-ST [36] BMN-BSN [37] FIGURE 2. A taxonomy of background subtraction-based moving object detection approach.…”

“…Overall, our method outperformed compared with the existing approaches. Fast-D -0.9961 0.0039 0.0286 0.5155 0.9714 0.9334 0.9520 SuBS [21] 2015 0.9962 0.0038 0.0405 0.5738 0.9594 0.9335 0.9462 SBS [17] 2019 0.9975 0.0024 0.0014 0.3862 0.9170 0.8672 0.8914 PBAS [19] 2012 0.9981 0.0019 0.1872 1.1500 0.8128 0.9585 0.8797 LOB [20] 2014 0.9980 0.0020 0.1893 1.1910 0.8107 0.9565 0.8776 ViBe [15] 2011 0.9939 0.0060 0.0018 0.7800 0.8950 0.7209 0.7986 SuperBE [18] 2019 0.9994 0.0005 0.0105 1.0844 0.5343 0.9552 0.6852 Table 5 shows the comparison of our proposed approach with the existing methods in terms of F1-score on CD-2012 and CD-2014 datasets. As we mentioned before, F1-score is the most important metric to evaluate the accuracy of BG/FG segmentation.…”

“…Later on, SACON's sample consensus policy was used by methods [15], [16] to improve classification accuracy. Recently, Chen et al [17] proposed sub-superpixel-based background subtraction (SBS) to increase detection accuracy. Additionally, in superpixel-level background estimation (SuperBE) [18], superpixel-based segmentation was proposed to reduce complexity.…”

“…However, to reduce the complexity, SuperBE considerably sacrificed segmentation accuracy. Though the algorithmic complexity of the above BGS methods [9]- [17] was low, the segmentation accuracy was very poor. Afterward, Hofmann et al [19] proposed a pixel-based adaptive segmenter (PBAS), newly created the dynamic background modeling policy, and used a sample consensus policy and gradient magnitude feature in addition to color feature.…”