3D-aided image augmentation for occluded pedestrian detection based on parallel vision
Songlin Bai, Quancheng Du, Yonglin Tian, Xiao Wang · 2023
Deep learning algorithms rely heavily on diverse and high-quality datasets to excel in complex tasks. However, in intricate traffic scenarios, manual data collection often falls short in terms of diversity due to the challenges of replicating real-world scenes. While techniques such as Generative Adversarial Networks (GANs) have been used for image generation, they frequently overlook the impact of object relationships on data diversity. In this paper, our primary focus is on the perception of pedestrians, especially those who are occluded, within complex driving scenes. We introduce a novel 3D-aided image augmentation technique known as "Drive-CP," which is based on the principles of parallel vision. This innovative approach enriches the dataset by generating instances of occluded pedestrians within the original images, thereby significantly enhancing the diversity of the training data. Experimental results consistently demonstrate that detectors trained on our augmented datasets outperform those relying solely on the original data. This work contributes to strengthening the robustness of deep learning algorithms in complex traffic scenarios, providing a more reliable foundation for practical applications.