A crowd-aware and self-supervised approach for object detection in wide area motion imagery

Date

2023-6-14

Author

Hatipoğlu, Poyraz Umut

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

120
views

76
downloads

Cite This

Detecting objects in Wide Area Motion Imagery (WAMI), an essential task for many practical applications, is particularly challenging in crowded scenes, such as areas with heavy traffic, since pixel resolutions of objects and ground sampling distance are highly compromised, and different factors disrupt visual signals. To address this challenge, we introduce two novel approaches including spatio-temporal cascaded architectures. Even though the deep detector architectures are similar, the training and operating principles of these two approaches differ greatly. While one of the approaches uses the given ground truth information and utilizes completely supervised learning principles, the second one does not need any prior knowledge about the locations of the targets to build a moving object detector. To the best of our knowledge, the latter approach is the first deep spatio-temporal approach trained without using ground truth information to detect objects in WAMI. Moreover, we propose a novel crowd-aware thresholded loss (CATLoss) function for training deep networks for detection in WAMI for improved performance in especially crowded areas. Furthermore, to incorporate more contextual information without introducing additional parameters, we extend prior networks used in the literature with dilated convolution layers. Overall, our approaches are causal, more generalizable, and more robust even in reduced spatial sizes. On the WPAFB-2009 dataset, we show that our solutions perform better than or on par with state-of-the-art without introducing any computational complexity during inference.

Subject Keywords

Moving object detection, Wide-area motion imagery, Convolutional neural networks, Deep learning, Self-supervised learning, Nonlinear Optimization

URI

https://hdl.handle.net/11511/104519

Collections

Graduate School of Natural and Applied Sciences, Thesis