Title: A hybrid optimisation enabled deep learning for object detection and multi-object tracking

Authors: J. Thirumalai; M. Gomathi; T.S. Sindhu; A. Senthil Kumar; R. Puviarasi

Addresses: Electronics and Communication Engineering, Prathyusha Engineering College, Thiruvallur, Tamil Nadu 602025, India ' Electronics and Communication Engineering, S.A. Engineering College, Thiruverkadu, Tamil Nadu 600077, India ' Electronics and Communication Engineering, C. Abdul Hakeem College of Engineering and Technology, Ranipettai, Tamil Nadu 632509, India ' Electronics and Communication Engineering, Kings Engineering College, Irungattukottai, Tamil Nadu 602117, India ' Electronics and Communication Engineering, Saveetha School of Engineering, Saveetha University, Sriperumbudur, Tamil Nadu 600124, India

Abstract: The potential of multi-object tracking (MOT) in academia and industry has drawn growing attention. Despite the various methods that have been suggested to address this issue, it continues to be difficult because of things like sudden changes in appearance and severe object occlusions. In this paper, a Jaya political search optimisation (Jaya-PSO) enabled ShuffleNet is developed for object detection (OD) and MOT. Initially, the input video is fed to video frame extraction. The extracted frames are fed into the object segmentation phase, where the segmentation is done by the mask-regional convolutional neural network (Mask-RCNN), trained by tangent squirrel search optimisation (TSSO). Here, TSSO is the integration of the tangent search algorithm (TSA) and squirrel search optimisation (SSO). Then, the object recognition is performed using ShuffleNet trained by Jaya-PSO, where the Jaya-PSO is from the Jaya algorithm, political optimiser (PO) and TSA. Finally, MOT is done by the Henry gas solubility optimised unscented Kalman filtering (HGSO-based UKF). The HGSO-based UKF is the integration of Henry gas solubility optimisation (HGSO) and unscented Kalman filtering (UKF). The measures utilised for analysis are accuracy, sensitivity, specificity and multiple object tracking precision (MOTP). The proposed method attained 92.9% accuracy, 92.1% sensitivity, 92.9% specificity, and 91.0% MOTP.

Keywords: object detection; multi-object tracking; mask-regional convolutional neural network; Mask-RCNN; tangent search algorithm; TSA; Jaya algorithm; political optimiser.

DOI: 10.1504/IJAHUC.2024.140033

International Journal of Ad Hoc and Ubiquitous Computing, 2024 Vol.46 No.3, pp.150 - 165

Received: 02 Aug 2023
Accepted: 04 Dec 2023
Published online: 15 Jul 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article