Enhancing surveillance in parking lots: a thermal infrared approach to human action recognition

To advance research in surveillance and Human Action Recognition (HAR) in challenging low-light environments, we construct a novel dataset: the Parking Lots Thermal HAR (PL-THAR). This dataset, captured using a long-wave infrared camera, serves as a benchmark for all-day surveillance, especially in outdoor parking lots with poor night-time lighting. To address the inherent blurriness of Thermal Infrared (TIR) images, we propose a novel HAR network based on thermal imaging, named THAR-Net. The network incorporates a multi-channel coordinated attention module, deformable convolutional networks, and cross-stage feature fusion modules to enhance feature extraction capabilities while reducing computational overhead. Comparative experiments with YOLOv8 show that THAR-Net achieves a 41.6% to 50.5% reduction in the number of parameters and a 26.2% to 33.1% decrease in floating-point operations, all while maintaining a detection accuracy of approximately 96%. Ablation experiments further validate the rationality and effectiveness of our network design. In addition, we present a new Detail and Contrast Enhancement (DCE) algorithm specifically designed for addressing the blurred edge problems in TIR images, thus enhancing human surveillance performance through thermal vision.