姿势自适应权重更新策略及尺度回归的煤矿井下目标跟踪算法

Attitude Adaptive Weight Updating Strategy and Scale Regression Algorithm for Target Tracking in Coal Mine

随着计算机技术的飞速发展,煤矿井下智能工作面的建设取得了巨大进步。目标跟踪作为井下智能工作面建设的一个重要环节,在保证作业人员安全,提升开采效率等方面发挥着重要作用。但是煤矿井下光照条件较差,当目标有较大外观变化和遮挡时,算法的跟踪预测框将不能准确跟踪目标。为了解决上述问题,设计出姿势自适应权重更新策略,该策略能计算出不同帧与模板帧之间目标外观姿势的变化程度,该变化程度作为目标估计网络调制向量更新的依据,动态地更新调制向量。通过该更新策略,使得算法对于目标尺度变化较大物体的跟踪精度得到提升。同时设计的中心点尺度回归损失函数,将中心点距离和高宽损失引入损失函数,使得目标框回归更加精准。

The rapid development of computer technology has led to great advances in the construction of smart underground work faces in coal mines.As an important element in the construction of intelligent underground working faces,target tracking plays a significant role in ensuring operator safety and improving mining performance.Poor lighting conditions in underground coal mines,however,mean that the model′s tracking prediction framework will be unable to accurately track the target when it has a large change in appearance and is obscured from view.To address the above issues,devise a pose adaptive weight update strategy,which can compute the degree of pose change of the target appearance between different frames and template frames,and this degree of change is used as the basis for updating the modulation vector of the target estimation network,and dynamically update the modulation vector.With this updating strategy,the model is able to improve tracking accuracy for objects with large changes in target scale.The designed centroid scale regression loss function simultaneously introduces the centroid distance as well as the height and width loss into the loss function,making the regression of the target image more accurate.