融合可变形卷积网络的鱼眼图像目标检测

Object Detection in Fisheye Images Combining Deformable Convolutional Networks

环视鱼眼图像具有目标形变大和图像失真的缺点,导致传统网络结构在对鱼眼图像进行目标检测时效果不佳。为解决环视鱼眼图像中由于目标几何畸变而导致的目标检测难度大的问题,提出一种基于可变形卷积网络的鱼眼图像目标检测方法。将Cascade_RCNN中固定的卷积层和池化层分别替换为可变形卷积层和可变形池化层,使用Resnet50网络提取候选区域以获得检测框,级联具有不同IoU阈值的检测网络进行检测框抑制。在公开鱼眼图像数据集SFU_VOC_360和本文所采集的真实道路场景鱼眼图像数据集上进行实验,结果表明,该方法在鱼眼图像目标检测中具有有效性,目标检测准确率高于Cascade_RCNN网络。

Due to the significant target shape changes and distortions in bird’s eye view fisheye images,the conventional network structures do not perform well in target detection for fisheye images.To address the geometric distortions of the target,which increases the difficulty of target detection in bird’s eye view fisheye images,this paper proposes an object detection method in fisheye images based on deformable convolutional network.The fixed convolution layer and pooling layer in Cascade_RCNN is replaced by the deformable convolution layer and deformable pooling layer.Then the candidate region is extracted by using Resnet50 to obtain the detection box,which is suppressed by cascading the detection networks with different Intersection-over-Union(IoU)thresholds.Experiments are carried out on the open fisheye image dataset SFU_VOC_360 and the manually collected fisheye image dataset of real on-road driving scenes.The experimental results demonstrate the effectiveness of the proposed method for object detection in fisheye images.Its detection accuracy is higher than that of Cascade_RCNN.