Research On Object Detection Technology With Stereo Vision Based On Deep Learning

In recent years,the application of object detection technology in people’s lives has become more and more widespread.However,the accuracy of the current object detection algorithms is not ideal,especially the detection accuracy of small and occluded objects.Most object detection algorithms only use one image from stereo cameras for object detection,while many devices in our daily life are equipped with stereo or more cameras,such as face recognition machines,sweeping robots,automatic driving systems,etc.Compared with single images,stereo images record images of objects from another perspective.Stereo images contain more information about the same object,so the object detection algorithms can learn more features of the object from stereo images,improving the accuracy of object detection.Based on stereo object detection algorithms,this thesis designs an effective stereo image feature fusion network.It constructs an algorithm for object detection using stereo images to improve object detection accuracy.The main work of this thesis is as follows.1.Proposed a stereo object detection algorithm,Faster R-CNN-S based on Faster R-CNN.This model is mainly based on the disparity attention mechanism,which aligns the right-view features to the left-view features to fuse stereo image features.Based on the KITTI benchmark dataset,Faster R-CNN-S improves the small object detection accuracy by 6.7%,medium object detection accuracy by 5%,and large object detection accuracy by 1.5% when compared with the Faster R-CNN.2.Proposed two stereo object detection algorithms,Deformable DETR-CA and Deformable DETR-D based on Deformable DETR.Deformable DETR-CA fuse stereo features based on convolutional attention modules,and Deformable DETR-D fuse stereo features based on deformable attention mechanisms.Compared with the Deformable DETR algorithm based on the KITTI benchmark dataset,the small object detection accuracy of Deformable DETR-CA network increased by 3.7%,the medium object detection accuracy increased by 1.9%,and the large object detection accuracy increased by 1.2%.The small object detection accuracy of Deformable DETR-D network increased by 3.4%,the medium object detection accuracy increased by 2.2%,and the large object detection accuracy increased by 0.8%.3.Designed and implemented a stereo object detection system,conducted a comprehensive requirements analysis,detailed the architecture and process design,and explained the various modules of the system.The above experiments show that these three stereo object detection algorithms proposed in this thesis have improved the accuracy of object detection on different evaluation indicators,showing that using stereo vision can effectively improve the accuracy of object detection algorithms.By applying stereo object detection algorithms to many devices in our daily life that are already equipped with stereo cameras,the practical effect of object detection can be improved at a low cost.