Research On Moving Object Tracking Algorithm In Video And Implementation Of Hardware Acceleration

Although the object tracking field has achieved good results over the years,it is still a challenging issue.With the development of artificial intelligence and deep learning,the object tracking algorithm is designed more and more complicated and time-consuming.It is also an urgent task to use heterogeneous platforms for quickly extracting features and parallelizing algorithms.An optimized searching strategy OSKCF algorithm for Kernelized Correlation Filters tracking is proposed to improve its real-time performance.The center position of target in a randomly selected video frame was detected,and the mean value and standard deviation of the target image patch were calculated respectively.Then a sort queue and two adaptive thresholds were set to discard unsuitable patches with certain features that differ greatly from the target patch.It was realized that the candidate patch in the next frame was detected and calculated with priority because whose mean values and standard deviations were both within a certain margin of the target patch.Then the KCF was extended from a single scale to 5 and 7 isobath scales.The test showed that when selected 5 levels of the scale factor,the tracking accuracy and speed were better.Experimental results show that the frame rate of the proposed algorithm increases about 10% than that of the original KCF algorithm,and the tracking accuracy increases about 2.2%,14.4%,24.9% than other algorithms such as KCF,CSK and Struct.In terms of hardware acceleration,a parallel scheme based on two different heterogeneous platforms was designed for the OSKCF algorithm.Including target initialization,cross-correlation matrix,calculation of regression coefficient and response maximum,and image feature extraction.Then used two parallelized algorithms to accelerate the following modules: FFT calculation,Gaussian kernel correlation calculation,image block search after optimizing strategy.Finally,measured the time consumption,power consumption and tracking accuracy of the two parallel algorithms,and compared performance with CPU serial algorithm.For fast Fourier transform and Gaussian kernel correlation,4.1～45.0 times of the acceleration ratio can be obtained when use the gray feature,and 4.0～38.3 times of the acceleration ratio can be obtained when use the FHOG feature.The time consumption of the two features on GPU side is controlled within 1ms.The acceleration ratio of the two features on Open CL-based FPGA platform is about 10 times higher than that of GPU,and the power consumption is also controlled within 1/6 of the GPU.In terms of the overall frame rate of the OSKCF algorithm,it obtains 9.0 times of the average speed ratio,and the average frame rate of the algorithm exceeds 500 frames,which achieves high speed real-time requirements and has good engineering value.