Object detection in automatic driving scenarios based on Semi-supervised learning

Due to the dramatic changes in the scale of the images captured in the autonomous driving scene and the complex and changeable environment, the detection is not small, and it is difficult to obtain a large number of labeled data images required for model training, while it is easier to obtain a large number of unlabeled data images. In order to solve the above two problems, a semi-supervised learning-based object detection model TransDet in autonomous driving scenarios is proposed. Firstly, a MSADark module with global attention is proposed in the feature extraction part to extract more global information of the image and capture long-range dependencies; secondly, in the feature fusion part, a positional attention weighted feature fusion network LAFFN is proposed for different feature fusion layers. Capture local location and channel information, enhance the ability of multi-level feature weighted fusion and network feature representation, and alleviate the impact of drastic changes in target scale; finally, a simple and efficient semi-supervised learning algorithm framework EODS is proposed. Further improved model performance. The experimental results show that the accuracy of mAP@50 increases from 55.1% to 61.6% when the improved model guarantees real-time performance, which is 6.5% higher than the accuracy of the latest YOLOv5 model. While ensuring the real-time detection speed, the model detection is improved. performance. Especially when using only a small amount of unlabeled data, the semi-supervised learning algorithm EODS improves the mAP.50 performance to 65.4%, and the improvement reaches 10.3%, which shows the effectiveness of the model in object detection in autonomous driving scenarios.