Human Instance Segmentation Based on Omega- Shape Using Deep Learning

Authors : Huma Sheraz, Zuhaib Ahmed Khan, Muhammad Awais

Pages : 31-41

Doi:10.53525/jster.1469697

View : 58 | Download : 101

Publication Date : 2025-06-28

Article Type : Research Paper

Abstract :Human detection and segmentation in an unconstrained environment is a very important and difficult task, having many important applications including tracking human beings, pedestrian detection, head count etc. Human detection in a single object environment is quite easy, but the problem becomes quite cumbersome in a crowded and cluttered environment. Most of the existing algorithms work on the detection and segmentation of the whole person in a scene. However, the performance of such algorithms degrades in case of occlusion and cluttered environment. To increase the performance in such an environment a technique is available that detects a distinct part of the human body which is “omega shape” instead of the full human body. However, the detection of bounding box also includes background pixels which limit the performance of the high-end applications such as tracking. Therefore, the objective of this research is to accurately segment the omega shape, so that the high-end applications have no background clutter in the human appearance model. We have trained and evaluated Mask R-CNN and YOLO+UNET and got a trade-off between accuracy and computation cost. The testing accuracy of Mask R- CNN and YOLO+UNET is 92.6% and 88.4%, while the computation cost is 6fps and 29fps, respectively.
Keywords : Deep Learning, Omega-shape, Machine Learning, Pedestrian Detection, Instance Segmentation.

ORIGINAL ARTICLE URL