AIGDet: Altitude-Information-Guided Vehicle Target Detection in UAV-Based Images

Abstract

Conducting extensive vehicle detection through the high-altitude perspective offered by unmanned aerial vehicles (UAVs) poses significant challenges. The high-altitude operation of UAVs to acquire a broader reconnaissance view results in low-resolution and densely packed vehicle targets in the captured imagery, creating substantial difficulties for vehicle detection. To address this, we propose a vehicle detection network specifically designed for UAVs, incorporating an end-to-end network that takes scale consistency constraints into consideration. The cornerstone of our method is the dynamic feature refinement module (DFRM), designed to overcome the feature attenuation and limitations in utilizing high-level prior information common in traditional approaches. Initially, we developed an adaptive target suggestion module based on the prior characteristics of the targets and scenes, and the scale consistency hypothesis of similar vehicles at different UAV flying altitudes. This module optimizes the number and scale of anchors by introducing prior information, facilitating preliminary localization of small-scale imaging targets. Subsequently, we constructed a multilayer feature purification structure based on a feature pyramid network (FPN) to refine bounding boxes at each level with height prior, integrating additional contextual information. This approach allows us to utilize more contextual information for vehicle detection while enhancing localization accuracy through detailed height prior. Our application and evaluation on multiple open-source datasets with height labels demonstrate that our method, with minimal parameter introduction, achieves excellent mean average precision (mAP) value. This underscores the effectiveness of our approach in UAV-based vehicle detection.