Research And Application Of Hybrid Parallel Structure Smart Camera

Traditional machine vision systems face several challenges,including large size,high power consumption,high cost,complex system design,difficult development,and long deployment cycles,which impede their adoption and widespread application in production situations.To address these limitations,this thesis proposes a new hybrid parallel architecture for intelligent camera systems that offers small size,high integration,and flexible deployment.Then,we develop and implement a software and hardware system based on this architecture and apply this intelligent camera system into two applications to demonstrate its effectiveness in real applications.After conducting an investigation and analysis of existing intelligent camera system architectures,this thesis first proposes a hybrid parallel architecture for intelligent camera systems,which includes CPU0(Linux),CPU1(Bare Metal),and FPGA.The hardware design of this architecture integrates these three platforms organically,and allocates image acquisition,image processing,and data transmission modules to different architectures to take full advantage of each platform.On the software level,the system separates process control procedure and functional computing module.By applying software reuse technology,the specific functional computing module can be customize while the system process control procedure is fixed.Therefore,this approach simplifies the development process,improves system scalability,and increases application flexibility.Furthermore,the efficacy of the proposed framework is demonstrated by utilizing it in two specific applications,structured light measurement and flame detection:(1)Following an in-depth analysis of the multi-frequency heterodyne threedimensional measurement algorithm,a FPGA parallelization strategy is designed on the proposed hybrid parallel architecture platform to accelerate computation.Moreover,the algorithm is optimized further by several advanced techniques,including floating-point to fixed-point conversion and pipeline parallelization,to enhance its efficiency while guaranteeing the accuracy of the output point cloud and reducing hardware resource consumption.(2)The flame detection network has been specifically optimized for low power consumption,low computational capability,and low clock frequency,which are characteristics of the hybrid parallel architecture.These optimizations,including network structure adjustment and parameters quantization,ensure high computational efficiency while minimizing accuracy loss.In addition,interference from flame-like targets can be further eliminated by installing an infrared filter.