基于多阶段提议稀疏区域卷积网络的城市交通目标检测

Urban Traffic Object Detection Based on Multi-Stage Proposal Sparse R-CNN

针对城市交通场景多目标检测算法检测速度慢,检测精度低等问题,本文提出多阶段提议稀疏区域卷积网络算法(Multi-stage Proposal Sparse Region-based Convolutional Neural Network,MPS R-CNN).算法主要有以下特点:提出了一种多阶段提议框过滤更新机制,提高算法检测精度;提出了一种双向并联特征金字塔网络(Bidirectional Parallel Feature Pyramid Network,BPFPN),增强了模型的特征融合能力;针对城市交通场景目标检测问题引入了CopyPaste数据增强方法和CIoU损失函数.实验结果显示,MPS R-CNN算法在Urban Object Dataset数据集上mAP达到了77%,算法检测速度保持在37 fps,优于目前其他城市交通场景目标检测算法.

Aiming at the slow speed and low accuracy of multi-object detection algorithms in urban traffic scenes,this paper proposes a multi-stage proposal sparse region-based convolutional neural network algorithm(MPS R-CNN). The algorithm mainly has the following characteristics: a multi-stage proposal box filtering update mechanism is proposed to improve the detection accuracy of the algorithm;a bidirectional parallel feature pyramid network(BPFPN) is proposed to enhance the model feature fusion capability;for the problem of object detection in urban traffic scenes, the Copy-Paste data augmentation method and CIoU loss function are introduced. The experimental results show that the MPS R-CNN algorithm achieves 77% mAP on the urban object dataset, and the algorithm detection speed remains at 37fps, which is better than other current urban traffic object detection algorithms.