基于边缘计算的实时目标检测算法

Real-time target detection algorithm based on edge computing

为了解决目标检测算法——SSD算法存在模型参数过多、存储需求大、对小目标检测效果不理想、难以应用于边缘计算设备的问题,提出一种基于SSD的改进算法SFF-SSD。该算法将SSD骨干网络改为ShuffleNet V2,并将网络中的Stage2_3和Stage3_7进行特征融合,形成新的特征提取层N_Stage3_7进行目标检测。同时,使用焦点损失函数处理不均衡正负样本的方法,改善SSD算法检测精度不稳定的问题。最后,采用对不同特征通道的重要性进行重标定的方式,引入SENet,提高网络的表示能力。实验结果表明:SFF-SSD算法在VOC2007数据集上参数量为5.465 MB,检测速度为77 fps,检测精度为73.96%,因此该研究方法可以在保证实时性和检测精度的同时满足边缘设备部署的要求。

In order to solve problem that target detection algorithm single shot multi-box detector(SSD)has too many model parameters and large storage requirements,the detection effect of small targets is not ideal,and it is difficult to apply to edge computing devices.An improved algorithm SFF-SSD is proposed based on SSD.The SFF-SSD algorithm changes the SSD backbone network to ShuffleNet V2 and performs feature fusion on stage2_3 and stage3_7 in the network to form a new feature extraction layer N_stage3_7 for target detection.At the same time,the method using focal loss function to process unbalanced positive and negative samples improves the problem of unstable detection precision of the SSD algorithm.Finally,using the method of re-calibrating importance of different characteristic channels,and introducing squeeze-and-excitation network(SENet)to improve the representation ability of the network.Experimental results show that the parameters amount of the SFF-SSD algorithm on the VOC2007 dataset are 5.465 MB,the detection speed is 77 fps,and the detection precision is 73.96%.Therefore,the research method can meet the requirements of edge device deployment while ensuring real-time performance and detection precision.