The Research On Tracking Algorithm Based On The Deformable Convolutional Neural Network

Computer vision is a complex information processing task whose main purpose is to perceive what exists in the surrounding environment and its exact location.Object tracking is an important basic research topic of computer vision.It has extensive research and applications value in the fields of intelligent video surveillance,video-based human-computer interaction,intelligent transportation system,intelligent visual navigation and 3D reconstruction.The rapid movement of objects,illumination changes,low resolution,occlusion,deformation,background interference,etc.are all challenges for this topic.Usually the online tracker is divided into four components: feature extraction,motion model,appearance model,and online update mechanism.The feature extraction of the object plays a decisive role in the decision-making judgment of the observation model.Although the convolutional neural networks are revolutionizing the object tracking subject because of their powerful feature representation capabilities.However,it is easy to track failure when the geometric deformation and scale are large or there are similar objects in the background.In response to the above problems,our work is based on the MDNet tracker to improve,the specific work is as follows:1.A deformable convolutional neural network algorithm tracker is proposed.Features extraction based on convolutional networks is now the implementation of mainstream trackers.However,due to the geometric invariance of the traditional convolution kernel,the receptive field is only a rectangular block of a fixed size.When the deformation or scale change of the tracking object is large,the tracker is less robust.Therefore,based on the MDNet algorithm,a geometrically deformable convolution operator is proposed,when the object deformation or scale changes greatly,the tracking is more robust.The deformable convolutional network feature is an important part of the tracking process,and has the advantages of a larger receptive field,higher reliability,and greater robustness to deformation scale changes.The focal loss,as an improvement of the cross entropy loss function,mainly solve the problem of imbalance between positive and negative training data in tracking.Therefore,a deformableconvolutional neural network algorithm tracker is proposed,which has achieved significant improvement in the OTB100 sequence with large deformation or large scale variation.2.A channel attention deformation algorithm tracker is proposed.In the framework of the DCT tracker,deformation and scale change problems have been solved.However there are similar objects in the background,the similar object score is close to the target causing tracking failure.The channel information mainly conveys the message “what is” identification information.The channel attention deformation network pays more attention to the key information,and weakens the background to highlight the foreground target,so that similar objects can be distinguished.Therefore,a channel attention deformation algorithm tracker is proposed,the background is weakened to highlight the tracking target.The group normalization is used to reduce the deviation between the training data and the verification data volume.The proposed algorithm shortens the pre-training time and improves the success rate and accuracy of tracking.