| With the continuous development of computer vision,a large number of application scenarios for video image preprocessing have been explored,and the demand for application landing is also increasing.Although the emergence of heterogeneous hardware platforms and hardware acceleration concepts has made it possible to apply target tracking to low-cost industrial applications on embedded devices,devices using different hardware architectures have increased development difficulties,and there are also certain limitations in the use of hardware resources and computing performance per unit of energy consumption,increasing the difficulty of project implementation.The content of this thesis is to study reliable target tracking algorithms that can be applied in industry,as well as their fast and stable implementation methods on heterogeneous embedded platforms.The research is mainly conducted from the following two aspects:(1)In terms of algorithm design,this paper studies target tracking algorithms under the traditional framework.While the important indicators of target tracking algorithms,such as strong real-time performance,high accuracy,and low resource utilization,cannot be met simultaneously,an improved kernel related target tracking algorithm is proposed to balance the requirements of the three indicators.The improved algorithm uses a Gaussian kernel function to process the features of the image,uses a kernel correlation filter to calculate the feature correlation response values for multiple scales of adjacent frames,and finally judges the target location and updates the filter parameters based on the response values.Experiments on OTB2015 dataset show that the improved algorithm has high accuracy,real-time performance,and small resource consumption.(2)In terms of the implementation of heterogeneous embedded devices,this paper designs a new heterogeneous programming framework after considering the hardware development time,development difficulty,and algorithm reuse rate.The framework transforms the programming process of the algorithm into model design,and then uses code generation tools to convert the model into hardware code.Although this approach increases the difficulty of timing design,it avoids direct programming and improves readability.Under the designed programming framework,this thesis builds an algorithm model based on the improved target tracking algorithm,and deploys the code generated by the model on the heterogeneous embedded device ZCU104 based on FPGA+ARM architecture.After continuous debugging and optimization,this thesis successfully completed the heterogeneous embedded hardware implementation of the design algorithm,achieving the goal of fast and stable implementation of the algorithm that can be applied in industry. |
