Object detection based on a robust and accurate statistical multi-point-pair model

Robust Object Detection in Severe Imaging Conditions using Co-Occurrence Background Model

“…We compared our algorithm with three methods: (1) GMM, [18] a standardized method among independent pixel-wise models; (2) Sheikh's KDE, [15] a representative method among spatialdependent models, which is different from the original KDE that it employs KDE over the joint domain (location) and range (intensity) representation of image pixels; and (3) GAP, [16] which is a predecessor and has a homologous methodology with CP3. The parameters for GMM were set as defaults in OpenCV tool; Sheikh's KDE [15] was set according to the author's recommendations with the size of the model [26,26,26,21,31]; and in GAP, W G ¼ 20,…”

“…Its accurate characteristics make it operable under several challenging severe imaging conditions. Compared with our earlier work GAP, [16] the proposed method has the following advantages: (1) CP3 employs a unique parametrized statistical model to describe each pixel-pair's co-occurrence rather than a fixed global doublesided threshold for all pixel-pairs in GAP; and (2) CP3 derives a self-adaptive threshold for each target pixel to select better-quality supporting pixels rather than a predefined threshold in GAP. Compared with some state-of-the-art independent pixel-wise models or spatial-dependence models, such an accurate background model significantly enhance the robustness of object detection in severe imaging conditions, e.g., foggy scene, low-light and noise, sudden illumination change, and narrow dynamic range, which can be observed in experimental section.…”

“…To employ not only an independent statistic of a pixel but also the spatial correlation between pixels, a group of spatialdependence methods has been proposed. [13][14][15][16] Seki et al [13] proposed a co-occurrencebased block correlation method, according to which the object can be detected as coarse local blocks. Toyama et al [14] proposed a three layers algorithm in which Weiner filters were employed, and it used region and frame-level information to verify pixel-wise background model.…”

“…This framework is suitable for dynamic background (such as ripple), but cannot deal well with illumination change. Zhao et al [16] proposed a learning-based approach, which aims to find high steady difference between pixels to offset the illumination change. Nevertheless, the presupposed threshold W G of the intensity difference influences the sensitivity of detection, as the intensity difference increases, the model becomes less sensitive; without optimizing…”

Robust Tracking Using Particle Filter with a Hybrid Feature

“…To overcome this problem, we utilize a novel robust feature called Grayscale Arranging Pairs (GAP) [10], [11], which was originally proposed for background subtraction. Compared with other background subtraction methods, the GAP feature builds a more accurate and robust background model that is flexible enough to handle different sets of complex conditions.…”