SAPNet: A Lightweight and Efficient Saliency-Aware Path Network for Tiny Human Detection in Thermal Aerial Imagery

Abstract

In aerospace applications, human detection from aerial thermal imagery has become increasingly important for surveillance and search-and-rescue missions. Designing a sys-tem that is accurate, efficient, and hardware-friendly remains challenging, as existing approaches often struggle with reliable detection and are computationally heavy. To address this, we propose a lightweight Saliency-Aware Path Network (SAPNet) that retains salient thermal signatures while leveraging dynamic path allocation to manage computational resources. The network introduces two novel modules: the Saliency-Aware Path and Channel-Reweighting (SPCR) module, which acts as a content-aware feature extractor to preserve subtle thermal signatures while reducing spatial and computational complexity, and the S3-CSP module, which enhances salient information aggregation from fused multi-scale features in the neck region. Extensive experiments on the TinyPerson benchmark demonstrate the effectiveness and efficiency of the proposed method, achieving 0.404 mAP@0.5 with only 3.4M parameters and 20 GFLOPs, highlighting its suitability for resource-constrained aerospace platforms.