Research On Automatic Driving Visual Recognition Method Based On YOLOv3

In recent years,with the development of artificial intelligence,especially deep learning,autonomous driving,a technology that existed only in peoples’ imagination has become increasingly close to reality.The complete meaning of automatic driving is inseparable from the support of advanced hardware and software equipment.In particular when the vehicle is driving,the cognition of surrounding objects needs the support of excellent computer vision recognition methods.YOLOv3 is the aggregator in the deep learning-based one-stage target detection algorithm.It has the advantages of fast speed,low background error detection rate and strong generalization ability,which can meet the requirements of real-time detection in the driving process and is a good choice for automatic driving visual recognition method.Of course it also has some inherent problems of YOLO series algorithms,such as the low recognition performance in multi-target scenarios,problems with the accuracy of target positioning,and the insufficient physical examination of small-scale objects.The purpose of this study is to search for a high-performance real-time visual recognizer in actual driving environment.Therefore it is necessary to test on the driving scene data set,to understand the recognition characteristics of the recognizer.Based on this,we hope to improve its performance and the confidence of visual recognition in autonomous driving applications.The average accuracy of the original YOLOv3 recognizer on BDD100 K is 0.423,which is much lower than its average accuracy on the COCO data set.For this reason,this paper first analyzes the impact of the Anchor size on the recognition performance of YOLOv3,and selects a set of Anchors through K-means clustering.Then,the network structure of YOLOv3 was improved,a YOLO head was added and a network structure of 4 YOLO heads was constructed.The average detection accuracy on the BDD100 K test set was 5% higher than that of the network structure of 3 YOLO heads.On the basis of four YOLO heads,an unbalanced YOLO structure was constructed,and its average precision after 32 rounds of iteration has been improved by 1% compared with the symmetrical structure of four YOLO heads.According to the analysis,the unbalanced 4 YOLO structure recognizer has advantages in the whole,especially in the small scale target recognition performance.This is also in line with our expectations for improved performance of small scale target recognition.