Cross-domain Person Re-Identification in Low Resolution Non-visual Sensor Network

Person Re-identification (RE-ID) is a computer vision technique for detecting specific persons in image or video sequences, enabling the recognition of target persons in non-overlapping camera views. This paper focuses on person re-identification under the FOVs of two infrared thermopile array sensors, which are similar to the low-resolution infrared cameras. A convolutional-recurrent neural network is constructed for closed-set identification and can achieve a cross-validation average accuracy of 91.3%. To solve the more challenging open-set identification problem, auxiliary methods such as outlier detection and manual feature extraction have been combined with neural networks, which address the limitations of neural networks in processing unknown data, leading to an average recognition accuracy of 88.6% for known classes and 83.2% for unknown class. The experimental results have demonstrated that low-resolution infrared images can be used for cross-domain person re-identification and achieve a relatively good recognition accuracy.