改进Mask R-CNN的车辆检测算法

Improved algorithm of Mask R-CNN for vehicle detection

为提升在不同复杂场景下的车辆检测性能,提出一种基于改进Mask R-CNN的车辆检测算法.在算法的主干网络ResNet50中引入PSA极自注意力机制提升主干网络特征提取能力;在特征金字塔顶层网络中添加一个带有ECA注意力机制的分支与原分支进行特征融合,缓解顶层特征由于通道降维造成的信息损失.重新设计卷积检测头使得边框回归更为准确,并使用余弦退火算法和Soft-NMS算法来优化训练过程和后处理结果.实验结果表明,改进的Mask R-CNN车辆检测算法相比原Mask R-CNN算法在复杂场景下具有更高的检测精度,在CNRPark-EXT测试集中平均精确度提高3.8%,在更具挑战性的MiniPark测试集中平均精确度提高7.9%.

To enhance the vehicle detection performance in complex scenarios,an improved vehicle detection algorithm based on Mask R-CNN is proposed in this paper.In the backbone network,the improved vehicle detection algorithm introduce the polarized self-attention(PSA)mechanism to enhance the feature extraction capability of the ResNet50 network.Additionally,in the top-level network of the feature pyramid network add a branch with an efficient channel attention(ECA)mechanism for feature fusion with the original branch.This helps alleviate information loss caused by channel reduction in the top-level features.The design of the convolutional detection head is also revamped to achieve more accurate bounding box regression.Furthermore,the cosine annealing algorithm and Soft-NMS algorithm to optimize the training process and post-processing results.Experimenta l results demonstrate that the improved Mask R-CN N vehicle detection algorithm outperforms the original Mask R-CNN algorithm in complex scenarios.On the CNRPark-EXT test set,it achieves an average precision improvement of 3.8%.On the more challenging MiniPark test set,the average precision is improved by 7.9%.