Research On Automatic Driving Target Detection Based On Improved Mask R-CNN

Auto-driving target detection technology focuses on the system’s perception of the external environment.In practice,the external environment of auto-driving is complex and changeable.The external environment of auto-driving has a great impact on the perception of the external environment of auto-driving,and there are still a large number of hard-todistinguish negative samples,such as signs,street lights,trash cans,etc.How to maximize the objective in complex and changeable environment? Standard detection accuracy is an important problem faced by automatic driving target detection technology at present.In order to improve the accuracy of target detection,this thesis uses Mask R-CNN,a classical algorithm of target detection and segmentation,and uses BDD100 k,a data set dedicated to automatic driving target detection,to train and test.Based on the algorithm,the shallow feature map and the high-level feature map are fused to optimize the algorithm.Finally,the accuracy of Mask R-CNN algorithm in target recognition is improved.The main work is as follows:1.Mask R-CNN algorithm is studied,and Mask R-CNN is trained using PASCAL VOC2007 and COCO datasets.The superiority of Mask R-CNN is verified by comparative experiments.2.In order to make a trade-off between the detection accuracy and speed,by changing the number of ResNet network layers,the appropriate number of network layers is selected for training when time permits.3.Mask R-CNN algorithm is applied to BDD100 k data set for training test.The experimental results show that the detection accuracy of Mask R-CNN algorithm is higher than Faster R-CNN,and the average detection accuracy of mAP is improved by 0.81%.4.Due to real-time changes in the real environment,small size of pedestrians,traffic signs and other targets,or weather,etc.,the unsatisfactory photographs will have an impact on the effect of automatic driving target detection.So an improved Mask R-CNN(IMask R-CNN)method based on shallow feature and high-level feature fusion is proposed,and it is trained and tested on BDD100 k data set.Its mAP is 34.47%.Compared with Mask R-CNN algorithm,the average accuracy of target detection is improved by 0.86%.5.By changing the feature fusion method,the original "element-sum" fusion method is changed to "concat" method based on channel addition,which is recorded as IcMask R-CNN.After training and testing in PASCAL VOC2007,its accuracy is improved.Applying it to the BDD100 k data set for training test,we find that its accuracy is improved by 0.12% compared with the mAP of IMask R-CNN.