Implementation And Optimization Of TLD Target Tracking Algorithm Based On ARM Processor

In recent years,artificial intelligence has gained unprecedented attention and achieved huge development.Artificial intelligence is constantly changing the industrial production and lifestyle,bringing us great convenience and superior experience.As one of the most important technologies of artificial intelligence,computer vision is to enable machines to perceive and understand the surrounding environment as our human eyes do.At present,the embedded platform based on ARM processor is limited in computing.However,advanced target tracking algorithms with good performance often have the characteristics of high complexity,heavy calculation burden,and difficulty in real-time implementation.Therefore,the realization and acceleration of target tracking algorithm for ARM processor and improving the real-time performance of the system are of significance in research and application.This thesis firstly implements the TLD target tracking algorithm based on the ARM processor,and then accelerates and optimizes the algorithm based on understanding it,which to improve the real-time performance of the algorithm.On the one hand,the algorithm is improved to reduce the calculation and complexity.On the other hand,the computing resources of the processor are fully exploited,and hardware acceleration is used to acceleration and optimization of the algorithm.This thesis mainly carried out the following work:1.Implementation and time-consuming analysis of the TLD algorithm.The principle of the TLD algorithm and the details of the implementation are analyzed.,so do the resources of the RK3399 software and hardware system.The algorithm operating environment is set up and the TLD algorithm is migrated to the RK3399 platform.Combining with algorithm theory and actual operation results,this thesis systematically analyzes the complexity and time-consuming of the TLD algorithm,discusses the time-consuming part of the algorithm running and the main factors affecting the real-time performance of the algorithm,so as to improve the real-time performance of the algorithm.2.Research on ARM processor-based acceleration optimization technology.This Thesis discusses the hardware architecture and software execution model of multi-core processor in depth,and further studies three different levels of acceleration optimization technology: multi-core concurrent technology based on Open MP,data-level parallel technology based on NEON multimedia engine and multi-pipeline technology.After the in-depth study of the NEON engine,this thesis makes a detailed analysis of the process of accelerating and optimizing NCC calculations using NEON technology,which is the core part in the research.As a result,the specific technical solutions are provided for the accelerated optimization of algorithms.3.Accelerate and optimize each module of the TLD algorithm.For the detection module of the TLD algorithm,local detection is introduced to reduce the number of scanned image blocks.Open MP is used to implement multi-core parallel computations of the random fern classifier,and NEON acceleration-optimized NCC is used to accelerate the speed of nearest neighbor classifier template matching.An optical flow tracker based on FAST corner points and a GMS mechanism is used in the tracking module to improve the original tracker and achieve more robuster and faster tracking.An online template update mechanism is introduced in the learning module to control the growth of the number of online templates to ensure real-time tracking over a long period of time.The experimental results show that the TLD algorithm based on the RK3399 platform wins a speedup of more than 3 times after accelerated optimization through various methods,which meets the real-time requirements of the target tracking system.