Automatic Driving Target Detection Based on Transfer Learning and YOLOv5-BEs Algorithm

To address the limitations of current road target detection algorithms, including insufficient small target detection capability, slow speed, frequent misdetection and omission, and long training time, this paper proposes a high-precision target detection model integrating transfer learning and improved YOLOv5 algorithm to satisfy the high requirements of detection speed and accuracy in autonomous driving scenarios.The Efficient Channel Attention (ECA) attention mechanism is first added to the model in order to increase the accuracy and efficiency of the model by strengthening the attention to tiny target characteristics.Second, to increase the multi-scale fusion capacity and the underlying information of the feature map, the Weighted Bi-directional Feature Pyramid Network (BiFPN) is utilized in place of the Feature Pyramid Network+Path Aggregation Network (FPN+PAN).Meanwhile, Scalable Intersection over Union Loss (SIoU_Loss) is used instead of Complete Intersection over Union Loss (CIoU_Loss) to enhance the localization accuracy and further optimize the model training effect.This study also creates a framework for transfer learning that moves the YOLOv5-BEs' already-learned information from the source domain training dataset to the target domain dataset.This makes the model better at training on the small sample dataset.Empirical findings indicate that the suggested YOLOv5-BEs model performs better than current algorithms, improving the Mean Average Precision ([email protected])by 1.3%~17.8%,and the Frames Per Second (FPS) by 4.46%~68.96%;through the transfer learning mechanism, the model's mAP @0.5 metric further reaches 68.2%, which is a 3% improvement from before transfer learning.The study's findings will offer an effective detection technique in the area of target identification for automatic driving, which has some potential uses.