Parallel Implementation for Real Time Person Matching System

Abstract

Local Binary Pattern multi-scale covariance descriptor (LBP_MSCOV) has been proved to be robust for video surveillance applications such as person detection, tracking and re-identification. Matching technique has recently grown in interest. It can be used to design person detection, tracking and re-identification. However, the original version is difficult to execute in real time. It requires a large data set and complex operations. Parallel implementation is adopted to achieve real time constraints. In this paper, we propose an optimized parallel model of a person matching algorithm based on LBP_MSCOV. For this end, a high-level parallelization approach based on the exploration of task and data levels of parallelism is adopted. First, an initial model is defined using only task-level parallelism. Second, this model is validated and analyzed at a high level of abstraction. Using the communication and computation workload results, the potential bottlenecks of this model are then identified. Concurrent optimizations are then performed to propose an optimized parallel model with the best workload balance. Finally, this model is validated and prototyped using a dual-core ARM-Cortex-A9architecture achieving up to 20.21 fps processing performance.