融合VovNet网络和可变形卷积的非机动车辆检测

Non-motor Vehicle Detection Based on VovNet Network and Deformable Convolution

针对道路监控下因监控探头高度角度不同,目标非机动车辆存在不同形式的模糊形变问题且特征信息不足造成的漏检误检现象,提出了一种融合VovNet网络和可变形卷积的非机动车辆检测模型.使用一次聚类连接网络(VovNet)结合原网络特点提出的CSPVovNet替换原有的CSPDarknet主干网络进行特征的提取,增强了有效特征的复用,缓解因深层卷积造成的小目标物体特征信息进一步丢失的问题.将可变形卷积引入到不同的网络层替换传统卷积,在公共数据集Pascal VOC2007和自建非机动车辆数据集上分别训练测试,根据最终性能选择YOLOv5-C方案.改进后的网络选取EIoU_loss作为定位损失,通过消融实验验证得出最终改进对网络性能有所提升,最终的网络优化结果较原YOLOv5s网络mAP提升了4.14个百分点,对漏检误检现象很好的缓解.

To solve missing and false detection caused by different fuzzy deformations and insufficient features of target non-motor vehicles due to different heights and angles of detectors under road monitoring,this study proposes a nonmotor vehicle detection model based on one-shot aggregation(VovNet)network and deformable convolution.CSPVovNet proposed by the VovNet network combined with the characteristics of the original network is used to replace the original CSPDarknet backbone network for feature extraction.This enhances the reuse of effective features and alleviates the further loss of features of small target objects caused by deep convolution.Deformable convolution is introduced into different network layers to replace the traditional convolution.Training and testing are carried out on the public data set Pascal VOC2007 and the self-built non-motor vehicle data set,respectively.The YOLOv5-C scheme is selected according to the final performance.The improved network selects EIoU_loss as the location loss.The ablation experiment shows that the final improvement improves the network performance,with the final network optimization result being 4.14 percentage points higher than the original YOLOv5s network in terms of mAP,which thus effectively alleviates missing and false detection.