本文针对目标跟踪应用, 提出了基于 Siamese-FC跟踪网络的改进卷积网络 Siamese-MF, 意在更进一步提升跟踪速度和准确性, 满足目标跟踪的工程应用需求。对于跟踪网络, 考虑速度和精度的权衡, 减少计算量, 增加卷积特征的感受野是改进跟踪网络的速度和精度的方向。在卷积网络结构上面进行改进结构创新, 改进主要集中为两点: 1) 引入特征融合, 丰富特征; 2) 引入空洞卷积, 减少计算量的同时增强感受野。 Siamese-MF算法实现了对于复杂场景目标的实时准确跟踪, 在公开数据集 OTB上测试速度达到平均 76 f/s, 跟踪成功率的均值达到 0.44, 而跟踪稳定性的均值达到 0.61, 实时性、准确性和稳定性均提升, 满足目标实时跟踪应用。

In this paper, aiming at the application of target tracking, an improved convolutional network Siamese-MF (multi-feature Siamese networks) based on Siamese-FC (fully-convolutional Siamese networks) is proposed to fur-ther improve the tracking speed and accuracy to meet the requirements of target tracking in engineering applications. For tracking networks, considering the trade-off between speed and accuracy, reducing computational complexity and increasing the receptive field of convolution feature are the directions to improve the speed and accuracy of tracking networks. There are two main points to improve the structure of convolution network: 1) introducing feature fusion to enrich features; 2) introducing dilated convolution to reduce the amount of computation and enhance the field of perception. Siamese-MF algorithm achieves real-time and accurate tracking of targets in complex scenes. The average speed of testing on OTB of public data sets reaches 76 f/s, the average value of overlap reaches 0.44, and the average value of accuracy reaches 0.61. The real-time, accuracy and stability are improved to meet the requirement in real-time target tracking application.