Research On Weak Target Tracking Based On Deep Learning

Visual object tracking is an extremely attractive and challenging research direction in the field of computer vision.Driven by the development of computer software and hardware technology,image processing technology,artificial intelligence technology,and chip technology,it has been widely applied to intelligent surveillance,intelligent transportation,intelligent human-computer interaction,battlefield situation reconnaissance,unmanned drone navigation and other commercial,military,and aerospace fields.Although outstanding tracking algorithms continue to emerge,however,due to its own uncertain factors such as target deformation,rotation,and motion blur,as well as external complex factors such as application scene lighting changes,background interference,and occlusion,the establishment of tacking algorithms with a high-precision,robust and good real-time performance is still facing huge challenges.Especially in aerial images and infrared images,due to extremely small target size,excessively low resolution,and highly strong background interference,these factors make the tracking task more difficult.The dissertation mainly focuses on the difficult points of weak target tracking(1)the feature representation ability of the target is weak;(2)the target motion state is more diverse and the position possibility is more variable;(3)the real-time requirement is strict,from the algorithm design level and hardware implementation level employing deep learning methods.The main work and innovations of the dissertation are as follows:1.An anti-occlusion Correlation Filtering target tracking algorithm based on multilayer deep convolutional features is proposed.Aiming at the problem of insufficient manual feature characterization ability and huge cost of depth feature calculation,an anti-occlusion Correlation Filtering target tracking algorithm based on multi-layer deep convolutional features is proposed.Firstly,the shallow features covering the positioning information and the deep features reflecting the semantic information are brought together to construct the robust target feature representation.Secondly,different levels of location correlation filters and scale correlation filters are dedicated to establish a framework suitable for weak targets,adaptively learning and determining the maximum response position.The phased evaluation strategy is adopted to update and to restore the model to alleviate the problem of error accumulation in the case of occlusion.Finally,the pre-trained deep convolutional network is optimized to reduce the computational cost by decreasing the feature dimension.Experimental results show that the algorithm balances the tracking speed while improving the tracking accuracy,and can adapt to complex conditions such as target occlusion.2.A Cascaded Region Proposal Network target tracking algorithm based on visual attention is proposed.Aiming at the problem that the pre-trained deep convolutional network is difficult to be fully applicable to online tracking,a Cascaded Region Proposal Network target tracking algorithm based on visual attention is proposed.Firstly,a deeper network model is adopted,and the Channel-Interconnection-Spatial attention module is integrated to highlight effective information and suppresses invalid information,so that the network can learn more robust target feature representation.Secondly,the Deconvolution Adjust Bblock is built to fusion cross-layer features,equipped with a three-layer Cascaded Region Proposal Network to construct a hierarchical tracking framework suitable for weak targets.Finally,to improve the tracking speed,the foreground-background classification and the bounding box regression response maps are acquired by the depthwise separable cross-correlation calculation,and the regional screening strategy is optimized.Experimental results illustrate that the algorithm further improves the accuracy and robustness,and can better adapt to complex scenes such as target appearance changes and background similarity interference.3.A Localization Confidence Network target tracking algorithm based on feature Super-Resolution is proposed.Aiming at the problems of small size of weak targets,low resolution,strong background interference and more variable position possibility,a Localization Confidence Network target tracking algorithm based on feature Super-Resolution is proposed.Firstly,the Super-Resolution feature maps for targets with small size,low resolution,and strong background interference,are generated by feature SuperResolution Generative Adversarial Network,which directly form a robust feature representation.Secondly,a joint tracking framework of online classification network and Localization Confidence Network is constructed.It provides a rough bounding box to distinguish foreground-background through online network,and uses offline network that integrates high-level knowledge to obtain accurate bounding box.Finally,the deep location confidence network is offline intensive training,while the shallow-level classification network is trained and updated online,which reduces computational cost.Experimental results display that the algorithm achieves more accurate position,and the effect of weak targets is more significant.4.A heterogeneous multi-core and real-time target tracking system based on the PYNQ framework is proposed and arranged.In order to meet the needs of actual target tracking application scenarios,a heterogeneous multi-core and real-time target tracking system solution based on the PYNQ framework is proposed,which is deployed on the ZYNQ heterogeneous platform.Firstly,the proposed anti-occlusion correlation filtering target tracking algorithm based on multi-layer deep convolutional features is divided into hardware and software to complete the system-on-chip construction.Secondly,the feature extraction acceleration module and the correlation filtering response acceleration module are designed to optimize the calculation process,and they are exported as the accelerated IP cores.Finally,via Jupyter Notebooks in the PYNQ framework,the accelerated IP core can be used as a hardware coprocessor to achieve data interaction from the bottom to the top.Experimental results on the ZYNQ heterogeneous platform ZCU104 show that the tracking accuracy of the heterogeneous tracking system is good,and the average speed can reach 27.9 frames per second.Algorithms suitable for weak target tracking in this dissertation are proposed from three aspects(1)designing a deep network model with strong feature expression ability,high discrimination ability,and good robustness to characterize target features;(2)constructing a high-performance target tracking framework suitable for weak targets;(3)reducing the computational cost and improving the tracking speed.The theoretical analysis and experimental results show that the proposed algorithm can improve the performance of weak target tracking.At the hardware implementation level,this dissertation designs and implements a set of deep learning hardware acceleration schemes for target tracking tasks.Under the circumstance of considering the tracking robustness,the heterogeneous tracking system has high execution efficiency,good portability and engineering application value.