基于孪生卷积神经网络改进的目标跟踪算法

Improved Object Tracking Algorithm Based on Twin Convolutional Neural Network

论文基于Tensorflow深度学习框架,搭建了一个基于深度学习的跟踪模型。通过卷积神经网络提取特征,运用互卷积运算得到响应特征图。通过端到端的训练得到一个分类网络和一个回归网络。其中分类网络用于判断跟踪到的目标是否正确,回归网络用于得到跟精确的目标定位。在训练数据上,以开源的数据集为主,采集到的数据集为辅。对于没有标注的图像采用OpenCV结合算法进行初步标注,然后再由人工检查。论文使用数据集训练了一个通用的目标跟踪器,实现了对一般目标的跟踪,并评估本算法的性能。

This paper designs a new tracking framework model based on deep learning based on Tensorflow.The convolutional network is used to extract the features,and the crossconvolution is used to get the response map.The classification network is used to judge whether the tracked target is correct,and the classification network is used to obtain accurate target positioning.Open-source dataset are used as the main data,and the collected dataset as supplementary data.OpenCV is used to label the data that is unlabeled,and then checked manually,which can reduce the workload and time cost.The data set is used to train a general target tracker to track the general target and evaluate the performance of the algorithm.The data set is used to train a general target tracker,it realizes the tracking of general targets,and evaluates the performance of the algorithm.