基于改进YOLOv3和立体视觉的园区障碍物检测方法

Hu Dandan,Zhang Lisha,Zhang Zhongting

摘要:

为了解决无人驾驶障碍物检测在园区场景中准确率低、实时性不足等问题，提出一种基于改进YOLOv3（You Only Look Once）和立体视觉的障碍物检测方法：YOLOv3-CAMPUS。通过改进特征提取网络Darknet-53的结构减少前向推断时间，进而提升模型检测速度，通过增加特征融合尺度提升检测精度和目标定位能力；通过引入GIOU(Generalized Intersection over Union)改进目标定位损失函数，通过改进k-means算法降低初始聚类点造成的聚类偏差，进而提高模型检测精度；通过立体视觉相机获得预测边界框中心点的深度信息，确定障碍物与无人车的距离。实验结果表明，提出的方法较原模型在园区混合数据集（KITTI+PennFudanPed）上平均精度提升了4.19%，检测速度提升了5.1fps；在自建园区数据集（HD-Campus）上平均精度达到98.57%，均能满足实时性要求。

Abstract:

In order to solve problems of low accuracy and lack of real-time performance of unmanned vehicle obstacle detection in the campus scene, an obstacle detection method based on improved YOLOv3 (You Only Look Once) and stereo vision was proposed: YOLOv3-CAMPUS. The forward inference time was reduced and the model detection speed was faster by im-proving the structure of the feature extraction network Darknet-53. The detection accuracy and target location accuracy were improved by increasing the feature fusion scale. Meanwhile, by using GIOU (Generalized Intersection Over Union), the target location loss function was improved. Enhanced K-means algorithm could reduce the cluster deviation caused by the initial clustering point, then the model detection accuracy was ameliorated. In additional, the depth information of the predicted boundary frame’s center point was obtained by stereo vision camera. Then the distance between the obstacle and the unmanned vehicle could be measured. Experimental results show the proposed method increases the average accuracy by 4.19% and the detection speed increases by 5.1 fps compared with the original model on the campus mixed data set (KITTI+PennFudanPED). On the self-built campus data set（HD-Campus）, average accuracy could reach 98.57%, and it could satisfy the real-time requirements by using improved method.