为使YOLOv2算法在保证检测速度的同时进一步提高目标检测的精确率,在YOLOv2模型的基础上提出RF-YOLOv2新模型。该模型先将KITTI数据集经过聚类,选出最适合KITTI数据集的候选框个数和候选框尺寸;其次在网络结构的训练部分采用残差块结构增加卷积层,提取更符合目标的特征描述;最后在网络结构的检测部分引入特征金字塔网络,将不同尺寸大小的特征图进行融合,使得低层特征图也具有丰富的语义信息。实验结果表明,RF-YOLOv2模型能获得更深层的特征、能融合更多尺寸的目标信息,改善了目标检测过程中由实际道路场景复杂、目标外形和结构多变等特点导致的检测率不高问题,在保证算法实时性的条件下,提高了对目标检测的精确率,RF-YOLOv2模型对大目标检测效果更佳。

To ensure detection speed and further improve object detection accuracy, a new model RF-YOLOv2 is proposed on the basis of the YOLOv2 model. In this new model, the KITTI data set is first clustered to select the most suitable number and size of candidate boxes. Next, a residual block structure is used to increase the number of convolutional layers in the training part of the network structure. This increase helps the model to extract more strong features to better describe objects. Finally, a feature pyramid network is introduced in the detection part of the network structure, fusing the feature graphs with different sizes. This network allows even low-level feature graphs to capture rich semantic information. Experimental results show that the RF-YOLOv2 model can gain the deeper information about features and can integrate more object size information. These improvements alleviate significant problems in current models that lead to low detection rates when actual road scenes are complex or when objects vary in shape or structure. The proposed model also improves object detection accuracy in real time detection and achieves better results for large object detection.