PAHD-YOLOv5:Parallel Attention and Hybrid Dilated Convolution for Autonomous Driving Object Detection

Object detection in autonomous driving scenarios is a complicated task. As onboard cameras always move, the object scale varies violently, which makes object distinction and localization difficult. To solve the issues mentioned above, we propose PAHD-YOLOv5 by improving YOLOv5. Firstly, the backbone is reconstructed using hybrid dilated convolution combined with a cross-stage partial network, whose role is to extract the contour information of objects and maintain the resolution of feature maps. Secondly, we give the same priority to channel attention and spatial attention, causing the network to modulate the channel and space weights simultaneously, whose role is to focus on the interest region of feature maps. Our proposed PAHD-YOLOv5 improves YOLOv5s by 2.3 mAP and 2.89 mAP on KITTI and BSTLD benchmarks, and the inference speed has reached 159FPS and 124.9FPS. PAHD-YOLOv5 shows its outstanding advantages in speed and accuracy compared with other advanced detectors.