Smart monitoring: employing person re-identification to uncover suspicious behavior

Person re-identification (Re-ID) includes matching people across multiple cameras plays an essential role in security, and surveillance, tracking people. People’s identification with multiple cameras presents significant challenges, particularly when behavior changes due to factors such as clothing variations, lighting conditions, and other people that affect multiple camera identification. A robust deep learning approach is proposed that aims to enhance the reliability of the re-identification model and novel technique for detecting suspicious activity using a combination of Euclidean Distance (ED) and Lookup Table(LUT). This model is trained on a publicly available VIPER dataset. We have also created our dataset to verify the validation through multiple cameras. The proposed model detects people by fine-tuning the YOLOv8 model, for trimming out objects along with individuals and afterward performing feature extraction and attribute evaluation. The feature extraction is performed in two steps: firstly, hand-crafted features used such as LOMO, which employs the HSV color space, and secondly, deep features retrieved from pre-trained neural networks models such as VGG16 and ResNet-50. The triplet loss and clustering methods are implemented to re-identify people and enable efficient interaction across multiple cameras or locations. The proposed Lookup Table (LUT) is developed through generated triplet loss features and a provided attribute matrix. It assists in the detection of suspicious activities based on information provided by identified individuals. In this research, a novel technique for activity detection is developed that combination of Euclidean Distance (ED) and Lookup Table(LUT) can accurately identify suspicious and non-suspicious actions. People have been detected using a re-identification model and used five camera configurations. The model achieves 90% accuracy in identifying suspicious and non-suspicious actions.