Onboard Person Retrieval System With Model Compression: A Case Study on Nvidia Jetson Orin AGX

A person retrieval system (PRS) in video surveillance identifies an individual based on descriptive attributes, a task that employs several computationally intensive deep learning models. We implement and analyse a PRS for pre-recorded videos on a graphics processing unit (GPU) and Nvidia Jetson Orin AGX. This paper presents a new Person Attribute Recognition (PAR) architecture, CorPAR, using three backbone networks, ConvNext, ResNet-50, and EfficientNet-B0. It enhances the F1-score by 4.1% with ConvNeXT-Base, 1.63% with the ResNet, and by 8.07% with EfficientNet-B0, surpassing the performance of the state-of-the-art Weighted-PAR method. The proposed method uses model compression techniques like quantisation and pruning with L1 regularisation to assess their impact on person retrieval. The study reveals that the PRS utilising EfficientNet-B0, with 32-bit quantisation, achieves the best performance, delivering a throughput of 22 frames per second and a True Positive Rate of 71% on Nvidia Jetson Orin AGX matching the performance of a model implemented using GPU.