Research On FPGA-Based Video Object Detection Acceleration Algorithm For Unmanned Airborne Platforms

With the rapid development of UAV research in recent years,object detection algorithms implemented on unmanned airborne platforms are widely used in various fields such as land survey,agricultural operations,and disaster relief.Deep learning algorithms headed by the CNN have greatly improved the accuracy of the object detection algorithm for the unmanned airborne platforms.Unmanned airborne platforms often have strict requirements on real-time performance of the object detection algorithm and the overall power consumption of the system.FPGA,with low power consumption,consists of a large amount of parallel logic resources,which is sufficient for processing computationally intensive tasks.It is a suitable hardware accelerator for the unmanned airborne platform running object detection algorithms.Firstly,this thesis explores the fixed-point approximation approaches of CNN,aiming at reducing the computational cost of CNN model on unmanned airborne platform.This paper successfully achieved 8-bit fixed-point approximation of the lightweight CNN model SqueezeNet V1.1 by linear quantization and static linear quantization,sacrificing less than only 1% accuracy.Secondly,this thesis uses SIMD parallel computing as the basic idea,and designs an FPGA-based object detection acceleration algorithm including convolution layer acceleration algorithm and pooling layer acceleration algorithm.Besides,this paper proposes two convolutional layer acceleration algorithm optimization strategies including parameter sharing and skipping 0 activation based on the computational characteristics of convolutional layers and the structure of the FPGA.Finally,this thesis tests and analyzes the performance and power consumption of the object detection acceleration algorithm on the DE10-Nano unmanned airborne platform under different parameters and optimization strategies.The FPGA-based object detection acceleration algorithm proposed in this paper is capable of processing about 8 frames per second at the 2.5W power consumption.The equivalent operation speed is 5.95 GFLOPS.Compared with related researches,the proposed algorithm has significant advantages in energy efficiency,and it is one of the state-of-art algorithms on resource-limited FPGA platforms.The FPGA-based object detection acceleration algorithm proposed in this paper lays a good foundation for the low-latency,low-power consumption object detection tasks of the UAV platform.