New multimodal pipeline detects humans under forest canopy with 0.83 mAP

A new research paper from Nitik Jain and Mangal Kothari tackles the challenge of detecting humans under sparse forest-canopy occlusion using a multimodal approach. The proof-of-concept pipeline evaluates three complementary techniques: experimental LiDAR returns through vegetation (which showed limited penetration for object-level detection), visible-thermal image fusion via a multi-scale transform and sparse-representation framework to enhance human saliency, and synthetic-aperture image formation through Airborne Optical Sectioning (AOS) to suppress canopy clutter. The researchers fine-tuned a YOLOv5 detector on the Teledyne FLIR thermal dataset and tested it on both thermal and fused imagery. Results show that visible-thermal fusion improves target visibility in low-contrast scenes, while AOS enhances ground-plane detection in synthetic forest imagery. The fine-tuned YOLOv5 achieved a mean average precision of approximately 0.83 on the top three FLIR classes, providing an initial baseline for UAV-deployable systems.

These findings are particularly relevant for search-and-rescue operations and surveillance in forested areas, where traditional optical sensors are often blocked by foliage. The study highlights that while LiDAR alone struggles, combining thermal and visible data with synthetic aperture techniques yields significant improvements. However, the authors note the need for dedicated forest datasets and real-time multimodal integration to move beyond this proof-of-concept stage. For professionals in remote sensing and emergency response, this pipeline offers a promising direction for drones equipped with multiple sensors to autonomously locate individuals under tree cover, potentially reducing search times in critical scenarios.