An Adaptive Weighted Object Tracking Combining Continuous Convolution Operator And Temporal Regularization

Object tracking has always been one of the hotspots in the field of computer vision,and has been widely used in intelligent control(such as unmanned aerial vehicles,robots,etc.),humancomputer interaction,autonomous driving and other fields.In the field of visual tracking,the efficient expression of features is the key to robust tracking.It is observed that in correlation filtering tracking,different convolution layers express different features of the object.An adaptive weighted object tracking algorithm combining continuous convolution operator and time regularization is proposed.Aiming at the problem of inaccurate object positioning,a continuous convolution operator method is proposed to transform discrete position estimation into continuous position estimation,so as to make position positioning more accurate.Aiming at the problem that different convolution layers have different capability of feature expression,an adaptive weighting method is proposed,which uses correlation filtering algorithm to adaptively fuse multi-layer convolution features to weaken background interference and enhance feature expression.In order to avoid the over-fitting of the trained correlator filter and improve the tracking performance,a regularization method is proposed to deal with the tracking failure in the case of occlusion or sharp deformation of the object.First,because the resolution of features of different layers of convolutional network is different,the deeper the layers are,the smaller the feature graph will be.In order to be able to integrate features of various resolutions together.The discrete feature graph in time domain is converted to continuous feature graph in time domain by cubic spline interpolation function,and then the continuous convolution operator is used to filter the acquired continuous correlation filter and continuous feature graph to make the object position estimation more accurate.Second,layer depth in different convolution neural network expression of different aspects of the object feature,namely shallow characteristics more location information,while the deeper features with more semantic characteristics,so if we can combine them to feature expression and tracking,will be compared with that by using deep or shallow feature better tracking effect.Firstly,the deep convolutional network structure is used to extract multi-layer convolutional features,and the weight of each layer feature in feature fusion in the next frame is determined by calculating the size of relevant convolutional response,so as to highlight the dominant feature and make the object more distinguishable from the background or interferences.Then,the correlation filter obtained from different layers of training is used to carry out correlation operation with the extracted features to obtain the final response graph.The position of the maximum value in the response graph is the position and scale of the object.Finally,a new model is constructed by introducing time regularization into the spatial regularization correlation filter.Combined with the adaptive feature fusion algorithm,an adaptive weighted target tracking algorithm combining continuous convolution operator and time regularization is obtained.Object tracking is a long tracking process.Due to the correlation filter of multi-image training and learning,which focuses on the nearest sample of the current frame,the filter obtained by the training of the nearest inaccurate sample may encounter the problem of overfitting,and lead to the tracking failure in the complex case of occlusion.After the introduction of time regularization,the algorithm in this paper can not only reasonably approximate the effect of spatial regularization correlation filter on the tracker trained by multi-frame training samples with one frame training sample,but also greatly improve the tracking efficiency.Experimental results show that the proposed algorithm is robust and efficient in tracking light variations,scale variations,background clutter,object rotation,occlusion and complex environments.