Multiple object detection in surveillance video with domain adaptive incremental fast rcnn algorithm

Abstract

Object detection is one of the key steps in wide variety of surveillance-based applications. This paper introduces an incremental class based Fast RCNN multiple object detection method that can be apply in various domain of surveillance video analytics. The two major components of proposed method are class incremental learning component and domain adaptation component. Class incremental learning helps for incremental adding of object classes with existing object classes for Fast RCNN. Domain adaptation component part improves the detection even the object presents in cross domain. Here basic object classes can be trained initially and then can add new object class incrementally according to new class requirement or according to domain. Benchmark datasets COCO, PETS S2L1, CD2014, Visor dataset were used for training and testing. Experimental results were showed that the proposed approach performs better in video datasets that contains challenges such as illumination variation, presence of shadow, partial occlusion and dynamic background.