基于鱼眼图像的目标检测算法研究

Research on the Object Detection Algorithm Based on Fisheye Image

鱼眼图像在使用前一般会先进行畸变矫正,但畸变严重图像的矫正会降低图像质量。为了提高目标检测精度与速度,文章提出了一种利用单个下视鱼眼摄像头代替多个普通摄像头的目标检测方案。其采用特征金字塔结构检测多尺度物体,并结合下视鱼眼的旋转与畸变特性进行算法优化,直接在原始鱼眼图像上进行目标检测。通过构建下视鱼眼数据集并进行实验,结果显示,所提出的鱼眼目标检测模型不仅精度较高,而且还能在嵌入式设备上快速运行。

Fisheye image is usually corrected before it is used, but image correction with serious distortion will reduce image quality.In order to improve the accuracy and speed of object detection, this paper presented a method which adopts a single bottom-view fisheye camera instead of multiple common cameras. It uses the feature pyramid structure to detect multi-scale objects, and combines the rotation and distortion characteristics of the fisheye to optimize the algorithm and directly detects the objects from the original fisheye image. In this paper, a downside fisheye dataset was constructed and experimented. The experimental results show that the fisheye detection model has high accuracy and can run fast on embedded devices.