基于双向融合SSD的目标检测算法

Object detection algorithm based on bidirectional fusion SSD

由于无人驾驶目标检测技术中对于小目标检测精度差、检测速度慢,降低了无人驾驶技术的安全性,实时性差。文中提出一种双向融合的SSD目标检测算法。该算法主要从两方面对SSD算法进行改进:使用基于深度可分解卷积的MobileNet代替传统的VGG-16网络作为SSD的主干网络提取特征,减小模型体积,大大地降低了计算量,提高了检测速度。该模型选取特定特征层,每一特征层经过递归执行核心模块和净化模块后的双向特征融合,组成分叉特征金字塔结构,递归地从深层和浅层双向循环语义和空间信息,有助于提高小目标检测精度。采用Tensorflow深度学习框架设置网络进行实验,在PASCAL VOC和COCO数据集上的实验结果表明,该算法在检测精度与检测速度上均有显著的提高,为无人驾驶技术的安全性与实时性提供了更好的保障。

Due to the poor detection accuracy and slow detection speed for small objects in the object detection of unmanned driving technology,the safety and real⁃time performance of the unmanned driving technology are reduced.An object detection algorithm based on bidirectional fusion SSD(single shot MultiBox detector)is proposed.The SSD algorithm is improved in two aspects.The MobileNet based on deep decomposable convolution is used to replace the traditional VGG⁃16 network and taken as the backbone network of SSD to extract features,reduce the model volume,greatly reduce the amount of calculation and improve the detection speed.In the model,the specific feature layers are selected.Each feature layer recursively performs bidirectional feature fusion of core module and purification module to form a forked feature pyramid structure,which recursively circulates semantic and spatial information from deep and shallow layers,so as to improve the accuracy of small object detection.The deep learning framework Tensorflow is used to set up network to perform experiments.The experimental results on datasets PASCAL VOC and COCO show that the detection accuracy and detection speed of the algorithm are significantly improved,which provides a better guarantee for the safety and real⁃time performance of unmanned driving technology.