Benchmarking Pedestrian Attribute Recognition Systems for UAVs Using Locally Collected Dataset: A Case Study in Thailand

Unmanned Aerial Vehicles (UAVs) are increasingly pivotal in security surveillance and crowd monitoring. A key step in automating these missions is pedestrian attribute recognition (PAR). This paper explores three PAR concepts for potential future implementation on UAVs with limited resources. First, a PAR system using a locally collected dataset in Thailand is implemented. The tracklets of 12,185 pedestrian images from CCTV camera footage–a UAV-like data source–are processed using YOLOv8 and BoT-SORT. These tracklets are manually annotated with 53 distinct attribute values, forming our Thailand Pedestrian Attribute Dataset (TPAD). Then, five models of attribute classification–DenseNet, EfficientNet, ConvNeXt, MobileNet, and ShuffleNet–are applied. The ConvNeXt model achieved the highest mean accuracy (mA) of 94.03%, while DenseNet yielded relatively similar mA of 92.93% but with lower complexity. Hence, DenseNet is a promising PAR-model candidate in UAV applications. Secondly, models trained on the TPAD, UPAR, and Market1501 datasets are compared. While data diversity enhanced performance with unseen data, models trained on locally collected datasets achieved higher accuracy. Lastly, the similarities and differences across the three datasets are examined. It revealed that models performed better on attributes with larger percentages. Some attributes were location specific. Variations in the percentages of common dominant attributes affected prediction performance. These findings emphasize the importance of using locally collected datasets for developing effective PAR systems. The work establishes a benchmark for creating more accurate PAR systems in the future.