Background subtraction (BGS) is a fundamental preprocessing step in most video-based applications. Most BGS methods fail to handle dynamic unconstrained scenarios accurately. This is because of overreliance on statistical model. In this paper, we develop a novel non-parametric sample-based background subtraction method. First, the background sample set is initialized by a clean sample frame rather than the first frame. This can avoid introducing a ghost when the first frame contains foreground objects. Here, we utilize the Gaussian mixture model to validate whether a pixel at the location is clean or not and construct the initialization of background model. Second, for an actual scenario with diversified environmental conditions (e.g., illumination changes, dynamic background), we employ normalized color space and a scale invariant local ternary pattern operator to handle these variations. In the meantime, in order to achieve high detection accuracy in the unconstrained scenarios without requiring any scenario-specific parameter tuning, we employ the perception-inspired confidence interval to modify the threshold in the color space. Third, the hole filling approach is used to reduce noise which comes from false segmentation, fill the blank area in the foreground region and maintain the integrity of foreground object. Our experimental results indicate that the proposed approach is superior to several state-of-the-art methods in terms of F-score and kappa index.