A Dataset and A Lightweight Object Detection Network for Thermal Image-Based Home Surveillance

Due to stringent privacy protection and all-day workability requirements, thermal cameras are more suitable than visible light cameras for analyzing indoor home scenes. However, application of object detection methods with thermal images are hampered by inadequate labeled indoor images and challenges of lightweight implementations. To address this issue, we first create a dataset called THS-DATA with 2,633 images, containing 19,362 person and pet targets, for indoor home surveillance object detection(1). Then, we propose a lightweight object detection architecture called THS-YOLO, where coordinate attention modules and adaptively spatial feature fusion (ASFF) modules are added to pruned YOLOv5s. In addition, pretraining strategies with regular RGB images are discussed. Experiments with the created dataset validate the effectiveness of the proposed network architecture and pretraining method.