Video Deraining for Mutual Motion by Fast Bilateral Filtering on Spatiotemporal Features

In video analytics, out of all the known challenges, (a) adverse unconstrained environmental condition resulting low illumination/ contrast and (b) mutual motion between camera assembly and object of interest are definitely the major ones. The challenges become multi-fold when the scene of interest gets impacted by both the aforementioned simultaneously. The first part of the current work has proposed a novel hierarchical scene categorization methodology based on selected spatiotemporal features. The procedure of determination of feature hierarchy and the detailed description of step by step methodology of the same has been provided along with comprehensive results. The second part of the current work proposes a novel method of processing rainy videos in different mutual motion scenarios by dynamic bilateral filtering and deep auto-encoder on time-sliced video frames. The proposed idea exploited the strength of range-domain filtering of bilateral filter, saliency extraction of deep auto-encoder and spatiotemporal model of time-slicing. Quantitative results of extensive experiments showed the effectiveness of our proposed algorithm in different degree of environmental degradation. The proposal of the scene categorization not only classifies the scenes based on mutual motion between camera assembly and object of interest, but it also triggers and influences the algorithm of deraining dynamically by offering required modulation.