Object Detection in Complex Traffic Scenes Based on Environmental Perception Attention and Three-Scale Feature Fusion

With the recent advancements in automated driving technology, object detection algorithms that can effectively respond to complex and diverse road traffic scenarios are especially important for driving safety during real driving. In this context, we conduct an in-depth study on object detection algorithms for diverse scenarios in autonomous driving. For diverse and changing backgrounds and multi-scale targets, we propose environmental perception attention (EPA) and the three-scale fusion module (TSFM) to improve the accuracy of object detection algorithms in complex traffic scenes. Environmental perception attention effectively improves the model’s ability to perceive the object by modeling long-range information and inter-channel relationships to direct the model’s attention to important task-related regions and important features in the image. The three-scale fusion module mixes features from different scales while introducing low-level feature map information, enabling the model to take into account the features of objects at different scales. In our experiments, we apply the proposed method to the YOLOv8 model for validation. The results show that compared to the performance of the baseline model on the BDD100K automated driving domain dataset with diverse and complex backgrounds, the mAP@0.5 metric of the improved model is increased by 1.3%. This makes the YOLOv8 more accurate and effective for the detection of different objects in the scenario, and it can better adapt to the different traffic scenarios and environmental changes.