• Figure 7-7: Convolution diagram for feedback template
  • Ultra fast cnn based Hardware Computing Platform Concepts for adas visual Sensors and Evolutionary Mobile Robots




    Download 3,22 Mb.
    Pdf ko'rish
    bet50/81
    Sana16.05.2024
    Hajmi3,22 Mb.
    #238917
    1   ...   46   47   48   49   50   51   52   53   ...   81
    Bog'liq
    Alireza Fasih

    Figure 
    7-7: Convolution diagram for feedback template 
    According to the Equation 7-4, for implementing an integrator we need a register to 
    accumulate the input value with the previous value. In Equation 7-4 we assign the value 
    2
    to the ‘K’ factor.
    (7-5) 
    9
    2
    )
    *
    *
    (
    )
    1
    (
    )
    (
    I
    u
    T
    y
    T
    n
    x
    n
    x
    B
    A
    To apply the ‘K’ factor we can operate a 9 time shift to right. Depending on the ‘K’ factor we
    can increase the accuracy. There is a trade-
    off between accuracy and speed. For a small ‘K’
    factor, we need more iteration to get the result. On the other hand by this way we can 
    change the time scale.
    7.9
     
    Hardware and software integration 
    Hardware modules are significantly faster than software functions. But in some cases we 
    need more flexibility in our system to access to the data and resources. The only way to get 
    a high performance is hardware and software integration. Using a high bandwidth 
    communication path between hardware modules and processors can be very helpful. For 
    example, for changing the template in the convolution and as well in CNN we need a 


     
    80 
    mechanism for loading data and storing it in the internal registers. For convolution based 
    video pre-processing we can just load the matrix coefficient that can change the 
    functionality of the system, and in the case of CNN we must change two different templates 
    and bias values. We know that the Impulse C from CoDeveloper is based on the 
    Communication Sequential Processor programming model. And it can generate the entire 
    necessary internal signal for synchronization and communication between hardware 
    modules. But for communication and exchanging values between hardware modules and 
    processor, there are some methods such as Processor Local Bus (PLB) and Fast and 
    Simplest Link (FSL). If we need only communication between two given modules, FSL is 
    the best, but for communication between a master and many slave modules, PLB can be 
    helpful. For Video and Image processing we do not need any high processing over the 
    processor, and the only reason behind that is initializing, re-configuring and debugging the 
    system; see Refs [124-126]. 

    Download 3,22 Mb.
    1   ...   46   47   48   49   50   51   52   53   ...   81




    Download 3,22 Mb.
    Pdf ko'rish

    Bosh sahifa
    Aloqalar

        Bosh sahifa



    Ultra fast cnn based Hardware Computing Platform Concepts for adas visual Sensors and Evolutionary Mobile Robots

    Download 3,22 Mb.
    Pdf ko'rish