Adaptive occlusion object detection algorithm based on OL-IoU.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Machine Learning

Background:

• Autonomous driving technology requires high accuracy in detecting traffic targets, especially in occluded environments.
• Existing object detection algorithms struggle with accuracy and real-time performance in complex occlusion scenarios.
• Addressing these limitations is crucial for the advancement of intelligent transportation systems.

Purpose of the Study:

• To propose an improved object detection algorithm, OCC-YOLOX, to enhance accuracy and real-time performance for occluded targets in autonomous driving.
• To leverage adaptive deformable convolution, coordinate attention, Overlapping IoU (OL-IoU) loss, and Fast Spatial Pyramid Pooling (Fast SPP).
• To validate the algorithm's effectiveness on diverse datasets and complex occlusion scenes.

Main Methods:

• An improved YOLOX algorithm named OCC-YOLOX was developed, incorporating adaptive deformable convolution.
• A coordinate attention mechanism was integrated to enhance focus on occluded targets.
• Overlapping IoU (OL-IoU) loss and Fast Spatial Pyramid Pooling (Fast SPP) were introduced to improve accuracy and reduce computational complexity.

Main Results:

• OCC-YOLOX demonstrated improvements in accuracy (2.76%), recall (1.25%), and average precision (1.92%) on fused public datasets.
• The algorithm was validated on KITTI, CityPersons, and BDD100K datasets, along with self-collected occlusion data.
• Experimental results show OCC-YOLOX outperforms existing mainstream algorithms, particularly in complex occlusion scenarios.

Conclusions:

• OCC-YOLOX significantly enhances the accuracy and efficiency of object detection in challenging autonomous driving environments.
• The proposed methods effectively address the limitations of current algorithms in handling occluded targets.
• This research offers valuable insights for improving object detection in intelligent transportation systems.