GPU-based Real-Time Multiple Moving Objects Tracking using Intergrated Spatial Region Graph for Video Surveillance

This paper presents the integration of a proposed enhanced multi-object color tracking, Partitioned Region Matching (PRM), and Spatial Region Graph (adjacency graph) for real time multi-object tracking. The problem of real-time object tracking is addressed by employing feature-based tracking technique that focuses on the integration of color feature tracking in regions of interest, and motion estimator which directly exploits computation of the region-level motion vectors through Partitioned Region Matching (PRM) that is based on the presence of gradients and semantically identify them according to their energy and other motion parameters. The preprocessed information are then converted to a spatial region graph (SRG) which is used as a starting point of a Markov Random Field (MRF) process, where regions are merged according to their semantics. The execution of the proposed system using GPU (NVIDIA GeForce GT 740M) showed that the processing time is enhancement by an order of 64% compared with its execution using CPU, which enabled an efficient onboard processing as well as centralized real-time processing of surveillance data, images and videos. The proposed method maps perfectly onto GPU architecture and has been implemented using NVIDIA CUDA. Experimental results on GPU for a sequence of frames, each of 460x480 pixels, showed that the implementation on GPU is 64 times faster than on CPU and confirmed the ability to process approximately 62 frames/s satisfying the necessary requirements for the correct subsequent tracking and reaching real time performance that demonstrates the suitability of the proposed system for real-time video surveillance.