Research And Implementation Of Key Technologies For Multi-target Tracking System

Target tracking is the most widely used technology in the field of computer vision.It has made important contributions in navigation,surveillance,human-computer interaction,virtual reality and even military.It has great application prospects and commercial value and has always been highly valued.During the past 30 years of development,many excellent classical algorithms have emerged.But due to the complexity of the environment,the ambiguity of the target's motion,and the various disturbances that the target faces in the tracking process,there is still much room for improvement in robust,accuracy and real-time performance.Efficient and stable tracking algorithms are still highly chalenging scientific research goals.The Tracking-Learning-Detection(TLD)tracking method is a pioneering combination of a tracking module,a detection module,and an online learning module,enabling long-term tracking of any single target.And the tracking method based on the correlation filter is a research hotspot in the past two years.The correlation filterbased algorithm transforms the complex convolution operation into simple multiplications and additions in the frequency domain through the discrete Fourier transform,and at the same time simplifies the operation using the properties of the cyclic matrix.So it's accurate and fast.In this thesis,the above two tracking algorithms are deeply studied.According to the longterm tracking architecture of tracking,detection,and learning separation in TLD algorithm,the long-term tracking of the target is realized by combining the fast and stable tracking algorithm of Kernelized Correlation Filters(KCF).Then it extends to multi-target tracking scenarios.The main innovations and research contents of this thesis are as follows:(1)Firstly,the background,research significances and research status of the thesis were introduced,and the current problems and chalenges faced in the target tracking were analyzed.Introduced the basic idea and tracking process of TLD algorithm and KCF algorithm.(2)In order to combine the KCF algorithm with the TLD framework,a multi-scale self-adaptive improvement is made to the KCF algorithm so that it can adapt with the scale change of the target in the tracking process.Then use the FHOG feature to replace the original HOG feature and speed up the algorithm.Combining the improved KCF algorithm with the TLD framework,the CF-TLD algorithm is proposed for long-term tracking of targets.According to the characteristics of the KCF algorithm,the TLD tracking framework is improved.Using the relative similarity of the target,the learning parameters of the correlation filter are dynamically adjusted to reduce the error accumulation in the tracking process.The detection module is used to deal with the loss of target.The situation prevents the KCF algorithm from tracking down the wrong target.Use the actual sequence to verify the effectiveness of the algorithm.The CF-TLD algorithm was further extended to multitarget tracking scenarios and its feasibility was verified through experiments.(3)Implement the porting of KCF algorithm on DSP platform.As an excellent target tracking algorithm,KCF algorithm has not been widely used in practical engineering.In this thesis,the KCF algorithm is implemented by using the TMS320C6657 platform for porting.The process of porting is described in detail,and the algorithm is optimized for the DSP platform.Finally,the dual-core DSP is used to achieve simultaneous tracking of dual goals.Through the actual sequence to verify its correctness,the experiment proved that the transplanted algorithm can work normally.