The objective of UAV-based vehicle detection is to accurately locate and classify vehicles in remote sensing images. It contributes to the improvement of intelligent city traffic management and enhances public safety monitoring. However, UAV offers more finely detailed remote sensing images than orbital satellites, resulting in more complex backgrounds, which poses a significant challenge. To tackle this challenge, we created a vehicle detection dataset named UAV Vehicle, which was collected by UAV flying at altitudes ranging from 250 to 400 m. The dataset comprises more than 200 remote sensing images, each with a size of 8192 × 8192, encompassing diverse complex scenes including urban, rural, and construction sites. However, the excessive complexity of the background conditions results in an imbalance between the number of foreground and background instances, consequently impacting the accuracy of final detection and localization. To address this issue, we proposed the RVDNet, which extracts precise positioning information from multiple scales, resulting in a substantial improvement in the model's localization performance. We developed an MSPP module, which employs random pooling methods to extract features at various scales, serving as a regularization technique to mitigate overfitting. Additionally, we introduced an attention mechanism to enhance the model's ability to represent and infer object positions more accurately. Extensive experiments conducted on the UAV Vehicle dataset demonstrate the efficacy of the proposed MSPP module and the introduced attention mechanism.
RVDNet: Rotated Vehicle Detection Network with Mixed Spatial Pyramid Pooling for Accurate Localization
Smart Innovation, Systems and Technologies
International Conference on Artificial Intelligence and Communication Technology ; 2023 ; Shenzhen, China June 09, 2023 - June 11, 2023
Proceedings of International Conference on Artificial Intelligence and Communication Technologies (ICAICT 2023) ; Kapitel : 27 ; 303-316
2024-03-21
14 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
P3SNet: Parallel Pyramid Pooling Stereo Network
IEEE | 2023
|Vehicle Color Recognition With Spatial Pyramid Deep Learning
Online Contents | 2015
|