Human re-identification using siamese convolutional neural network on Nvidia Geforce RTX 2060

Abstract

Human reidentification in multiple cameras with disjoint views is to match a pair of humans appearing in different cameras with non-overlapping views. Human reidentification has been extensively studied in recent years because it plays a significant role in many applications such as human tracking and video retrieval. However, human re-identification is a challenging task due to varying factors such as color, pose, viewpoint, lighting conditions, low resolution and partial occlusion. Most of the existing methods in handling human re-identification task are based on various handcrafted features and metric learning. However, hand-crafted features method requires expert knowledge and requires a lot of time to tune the features and metric learning methods are not powerful enough to exploit the nonlinear relationship of samples. The main objective of this thesis is to implement Siamese Convolutional Neural Network (SCNN) for person re-identification task in multiple cameras on the NVIDIA® GeForce RTX™ 2060 platform, including person detection. This continuous with validation of the applicability of SCNN and compare with existing techniques. In this work, global and local features of human images are extracted from SCNN. The proposed SCNN consists of two identical Convolution Neural Networks with common parameters that can automatically learn hierarchical feature representations from image pixels directly which has advantages than the hand-crafted design and metric learning method. Experiments were conducted with CUHK02 offline database with non-overlapping cameras. The proposed technique demonstrated a person re-identification using SCNN on the NVIDIA® GeForce RTX™ 2060 platform.