Enhancing YOLO for occluded vehicle detection with grouped orthogonal attention and dense object repulsion

Jinpeng He, Huaixin Chen, Biyuan Liu, Sijie Luo, Jie LiuScientific 2024 · DOI

Abstract

In real-life complex traffic environments, vehicles are often occluded by extraneous background objects and other vehicles, leading to severe degradation of object detector performance. To address this issue, we propose a method named YOLO-OVD (YOLO for occluded vehicle detection) and a dataset for effectively handling vehicle occlusion in various scenarios. To highlight the model attention in unobstructed region of vehicles, we design a novel grouped orthogonal attention (GOA) module to achieve maximum information extraction between channels. We utilize grouping and channel shuffling to address the initialization and computational issues of original orthogonal filters, followed by spatial attention for enhancing spatial features in vehicle-visible regions. We introduce a CIoU-based repulsion term into the loss function to augment the network's localization accuracy in scenarios involving densely packed vehicles. Moreover, we explore the effect of the knowledge-based Laplacian Pyramid on the OVD performance, which contributes to fast convergence in training and ensures more detailed and comprehensive feature retention. We conduct extensive experiments on the established occluded vehicle detection dataset, which demonstrates that the proposed YOLO-OVD model significantly outperforms 14 representative object detectors. Notably, it achieves improvements of 4.7% in Precision, 3.6% in [email protected], and 1.9% in [email protected]:0.95 compared to the YOLOv5 baseline.

Introduction

Object detection has seen significant advancements, yet "occlusion" remains a primary hurdle in complex traffic scenarios. Occlusion typically occurs in two forms: inter-class occlusion, where non-vehicle objects like trees or buildings hide a car, and intra-class occlusion, where vehicles hide each other. Standard detectors often lose key features in the former case or fail to distinguish between tightly packed targets in the latter, leading to overlapping prediction boxes that are incorrectly discarded during post-processing. YOLO-OVD aims to bridge this gap by forcing the model to focus on visible vehicle fragments and better separate dense targets.

Methodology

The YOLO-OVD architecture introduces three primary innovations designed to handle high-occlusion environments:

  • Grouped Orthogonal Attention (GOA): This module uses orthogonal filters to maximize information extraction across different feature channels. By utilizing grouping and channel shuffling, the model maintains computational efficiency while ensuring it pays more attention to the "unobstructed" regions of a vehicle.
  • Laplacian-Guided Feature Fusion: The framework integrates a classical Laplacian Pyramid into the deep learning neck. This helps the network retain fine-grained spatial details that are often lost during standard downsampling, leading to faster training convergence and more detailed feature retention.
  • CIoU-Repulsion Loss: Standard loss functions are augmented with a "repulsion" term based on Complete Intersection over Union (CIoU). This term pushes prediction boxes away from neighboring ground-truth boxes they are not intended to match, which is crucial for accurately separating cars in dense traffic jams.

Dataset: UADETRA-OVD

To evaluate the model, the researchers created UADETRA-OVD, derived from the large-scale UA-DETRAC dataset. The team sampled 8,677 images, specifically filtering for images rich in occlusion and providing new manual annotations for vehicles that were previously unlabeled or ignored in the original set.

Key Results

In benchmarks against 14 state-of-the-art detectors, including YOLOv7 and YOLOv8, YOLO-OVD demonstrated superior performance. Compared to the YOLOv5 baseline, the model improved Precision by 4.7%, [email protected] by 3.6%, and [email protected]:0.95 by 1.9%.

Qualitative comparison of detection results (a-c) and activation maps (d-f) between YOLOv5, YOLOv7, and the proposed YOLO-OVD.

As shown in the qualitative analysis, activation maps (CAM++) reveal that YOLO-OVD has much sharper focus boundaries around visible vehicle parts compared to standard models, which tend to exhibit blurry or misplaced attention in occluded scenes. Despite these enhancements, the model remains efficient at 12.9M parameters, making it more lightweight than many competing architectures like Faster R-CNN or YOLOv7.

Conclusion

YOLO-OVD effectively addresses the challenges of vehicle detection in occluded urban environments through specialized attention mechanisms and repulsion loss. Future work will focus on adapting these mechanisms for drone-based surveillance, where small object detection and different occlusion patterns (such as overpasses) present the next major bottleneck.

Summary updated

Abstract

In real-life complex traffic environments, vehicles are often occluded by extraneous background objects and other vehicles, leading to severe degradation of object detector performance. To address this issue, we propose a method named YOLO-OVD (YOLO for occluded vehicle detection) and a dataset for effectively handling vehicle occlusion in various scenarios. To highlight the model attention in unobstructed region of vehicles, we design a novel grouped orthogonal attention (GOA) module to achieve maximum information extraction between channels. We utilize grouping and channel shuffling to address the initialization and computational issues of original orthogonal filters, followed by spatial attention for enhancing spatial features in vehicle-visible regions. We introduce a CIoU-based repulsion term into the loss function to augment the network's localization accuracy in scenarios involving densely packed vehicles. Moreover, we explore the effect of the knowledge-based Laplacian Pyramid on the OVD performance, which contributes to fast convergence in training and ensures more detailed and comprehensive feature retention. We conduct extensive experiments on the established occluded vehicle detection dataset, which demonstrates that the proposed YOLO-OVD model significantly outperforms 14 representative object detectors. Notably, it achieves improvements of 4.7% in Precision, 3.6% in [email protected], and 1.9% in [email protected]:0.95 compared to the YOLOv5 baseline.
BibTeX
@inproceedings{he2024enhancing,
  author = {Jinpeng He and Huaixin Chen and Biyuan Liu and Sijie Luo and Jie Liu},
  title = {{Enhancing YOLO for occluded vehicle detection with grouped orthogonal attention and dense object repulsion}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41598-024-70695-x},
}
Figure