An Improved Version of Texture-based Foreground Segmentation (accepted at ICCSCI’18)
Notes: This work is accepted for oral presentation at The 3rd International Conference on Computer Science and Computational Intelligence (ICCSCI’2018)
The publication is an open access paper, please visit the following link: [here]
Foreground segmentation is one of moving object detection techniques of computer vision applications. To date, modern moving object detection methods require complex background modeling and thresholds tuning to confront illumination changes. This work proposes an adaptive approach based on non-overlapping block texture representation. It aims to design a computationally light and efficient solution to improve the robustness of detection. We evaluate our proposed method on internal and public sequences and provide the quantitative and qualitative measurements. Experimental results show that the proposed method can improve the results of previous method and suitable for real-time challenges.
Moving object detection is an underlying task in computer vision field, especially for extracting the insights from the video surveillance footages. This technology is often implemented to a higher-level vision, such as the human gesture and gait recognition, vehicle counting, and intelligent robot vision. Recently, the object recognition technology, which is able to improve the safety of outdoor activities (e.g., vehicle, pedestrian, etc.), has been massively deployed on the Smart City solution in developing countries. Some kinds of literature have been comprehensively reviewed the classification of background subtraction models. You may read some survey papers, such as Bouwmans T. [1], etc.
In this current work, we present an adaptive block-based background subtraction method, inspired by PBAS[2] method. For simplicity, we name our approach Block-based Adaptive Segmenter (hereinafter referred to as BBAS). Instead of relying on pixel intensities, we use texture information and model the observed block by incorporating the neighbourhood information.
Our proposed model is based on the model proposed by Yeh et al.[3] which uses a non-overlapping block to subtract the background. The initial step is to divide the current frame into n x n non-overlapping blocks. It is typically set the n to 4, in most of the experiments. Finally, the process converts each block into a binary bitmap. When a new bitmap, BM-observed, comes in, it will be compared against the several number of weights BG model BM-model, where the total number of weights is set as parameter K in the algorithm. Afterwards, the similarity between BM-observed and BM-model is calculated through Hamming distance. If the distance between BM-observed and BM-model exceeds the predefined threshold, the block is classified as foreground (FG), and vice versa.
Block-based Adaptive Segmenter (BBAS)
Single-phase improvement
Our first improvement is to modify the update scheme of Yeh, et.al.s method[3]. Compared to this method which only updates the corresponding BG model of the observed block, our proposed idea also incorporates the adjacent blocks by prior checking the similarity and update the adjacent BG models, accordingly. More specifically, different from [2] that select and update randomly neighboring pixels, we first check the similarity of the observed BM-model with BM-adjacent model via hamming distance. If the similarity of models exceeds the TH-adjacent, then all BM-adjacent model is replaced by its current binary bitmap. In this research, the TH and TH-adjacent are set to 8 and 13, respectively. The benefit of using BBAS is that the fragment issues (in the inner region of a detected object) can be alleviated. This benefit can be seen by comparing the qualitative results of our dataset shown below. Moreover, uninteresting foreground region can also be corrected and regarded as BG region.
Final improvement
In Yeh, et.al.[3], the bit transition is calculated to determine the mode of a block. If the block is complex, then the higher level is employed (2 or 3-bits mode, instead of 1-bit mode). In our proposed scheme, the result from the first phase allows further improvement by utilizing bit transition in order to check the complexity of observed block and adjacent blocks. If the observed block is classified as BG, we then check the complexity of current adjacent blocks and whether it is FG block or not. As shown in the qualitative results of final proposed method below, our final contribution can correct the mislabel blocks inside the detected objects.
Experimental Results
We provide the comprehensive evaluation including qualitative and quantitative measurements. In this research, we use our own dataset and CDNet dataset[4]. Our own dataset shows a person moving slowly towards the stairs. Our final proposed method can solve the fragmented issues when applied on our dataset. Besides, our method is competitive to the previous method on CDNET2014[4] challenges (can be found on the full version of paper at ICCSCI’18). Here we tested 4 videos from baseline category. Note that, we only use CPU-based processing, that is suitable for thin client.
- Bouwmans, T.. Traditional and recent approaches in background modeling for foreground detection: An overview. Computer Science Review 2014;11–12:31–66.
- Hofmann, M., Tiefenbacher, P., Rigoll, G.. Background segmentation with feedback: The pixel-based adaptive segmenter. IEEE Computer
Society Conference on Computer Vision and Pattern Recognition Workshops 2012;:38–43 - Yeh, C.H., Lin, C.Y., Muchtar, K., Kang, L.W.. Real-time background modeling based on a multi-level texture description. Information
Sciences 2014;269:106–127. - Wang, Y., Jodoin, P.M., Porikli, F., Konrad, J., Benezeth, Y., Ishwar, P.. CDnet 2014: An expanded change detection benchmark dataset.
IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops 2014;:393–400.
