A multi-strategy integrated improved YOLOv5 model and its application in target detection

Deep learning techniques have the potential to significantly improve target detection speed and detection accuracy in autonomous driving. Most existing target detection algorithms have poor real-time performance and poor accuracy. Moreover, the algorithm is difficult to be deployed. Based on analysis of the existing problems of target detection in autonomous driving, this paper puts forward an improved YOLOv5 algorithm. First, the extract feature network of YOLOv5 is replaced with the FasterNet-T0 model to reduce model parameters. Next, in the fusion part (Neck) of the network, the attention mechanism CBAM is introduced to improve the accuracy of the target detection. Then, the Slim-Neck framework is presented to improve the computational efficiency of the model. Finally, considering that CIoU may be replaced by Inner-IoU to improve the accuracy and generalization of the model, in order to verify the effect of the improved model, this paper uses the improved YOLOv5 algorithm model to extract the specific categories from VOC2007 dataset and DOTA dataset. After experimentation. The results show that the model we made, the improved YOLOv5 algorithm model is capable of getting a 0.848 and 0.748mAP @ 0.5 of the VOC2007 dataset and DOTA dataset for the specific category extraction, which verifies its effectiveness. Compared with other algorithms, the improved YOLOv5s algorithm is more competitive. On the VOC2007 data set, mAP live/voc, 0.5 increased by 5.1% for a specific category acquisition, and mAP live/voc, 0.5:0.95 increased by 6.4% relative to the original YOLOv5s algorithm, and the number of parameters reduced by 39.2% from 7035811 to 4276490. On the DOTA dataset, mAP@0.5 increased by 0.9% for a specific category acquisition, and mAP@ 0.5:0.95 increased by 0.4%, which demonstrates that this method improves the accuracy and efficiency of vehicle detection to some extent. In this study, DOTA dataset was used as an experimental dataset to verify the performance of YOLOv5s in the category of vehicle detection. This study is based on the follow-up research of autonomous driving. It lays the foundation for the follow-up research of autonomous driving to get better real-time performance and accuracy of object detection.