Multiple Action Recognition in Compressed Domain Videos using Cognitive Learning

Abstract

Abstract newlineIn the realm of computer vision, multiple human action recognition in compressed video newlinepresents significant challenges. Human action recognition in videos finds significant applications newlinein surveillance, elderly person monitoring at home/hospital, human-computer newlineinterface, video summarization and many more such applications that require scene understanding newlineand human activity analysis. With the motivation provided by the recent newlineliterature on human action recognition in raw(pixel) domain or compressed domain considered newlineby many researchers in view of the various significant applications, the problem newlinestatement for the present research work is formulated as To design and develop efficient newlinemethods for multiple human action recognition in compressed video domain using newlinecognitive learning . In this thesis, several novel methodologies are proposed to enhance newlineaction recognition within compressed video domains, by considering foundational advancements newlinein multi-object tracking, and addressing limitations of existing techniques. newlineInitially, an advanced object detection algorithm is introduced by refining YOLOv3 newlineand integrating Deep SORT for real-time multi-object tracking. This approach, tested newlineon public and proprietary datasets, demonstrates superior accuracy and performance in newlinepedestrian tracking compared to traditional methods. Building upon multi-object tracking, newlinea focus is placed on cyclist detection and tracking. Using the modified YOLOv3 newlineand Deep SORT, the method effectively tracks cyclists and calculates their speed using newlineoptical flow. Evaluations on the KITTI and SCD datasets validate the efficacy of this newlineapproach in real-world scenarios. newline