Journal of Applied Optics, Volume. 43, Issue 4, 682(2022)

Implementation of real-time deep learning tracking algorithm based on NPU

Xi HE... Liangfu LI, Jiaoying WANG, Jie WANG, Xiaoyan LU, Jun QIAN, Yizhou YANG and Peizhen LIU |Show fewer author(s)
Author Affiliations
  • [in Chinese]
  • show less
    Figures & Tables(22)
    Parameters update during training and inference
    Comparison between classification score and IoU overlap score
    Same examples of IoU loss with different locations
    Loss judgement added into centric position
    Network before and after pruning
    Flow chart of channel pruning
    Structure diagram of algorithm
    Initialization unit
    Tracking unit
    Intermediate results of tracking unit
    Online update unit
    Function of online update unit
    System structure diagram of Atlas200 DK
    Test results of CarDark sequence
    Test results of CarScale sequence
    Test results of SUV sequence with occlusion
    Test results of SUV sequence after occlusion
    • Table 1. Comparison of performance parameters of important GPU and NPU platforms

      View table
      View in Article

      Table 1. Comparison of performance parameters of important GPU and NPU platforms

      参数TX2NVIDIA Xavier华为Atlas 200
      计算单元NVIDIA Pascal架构搭载 256个NVIDIA CUDA 核NVIDIA Volta架构搭载 384个NVIDIA CUDA核 和 48个张量核2 个达芬奇AI 核
      算力(fp16)/FLOPS1.3T6T10T
      算力(int8)/OPS21T22T
      功耗/W151512
    • Table 2. Target contrast and size of data set

      View table
      View in Article

      Table 2. Target contrast and size of data set

      视频序列图像分辨率/pixels最小对比度最大对比度最大尺寸/pixels最小尺寸/pixels
      BlurCar1640×4800.0210.059(142,128)(80,80)
      BlurCar2640×4800.0230.143(239,184)(89,81)
      BlurCar3640×4800.0160.026(98,83)(77,66)
      BlurCar4640×4800.0560.114(202,174)(143,121)
      Car1320×2400.0030.047(66,55)(17,15)
      Car2320×2400.0230.043(64,52)(46,39)
      Car4320×2400.0300.122(110,85)(52,44)
      Car24320×2400.0080.037(56,51)(27,24)
      CarDark320×2400.0080.013(35,30)(29,22)
      CarScale640×2720.0050.216(306,123)(44,21)
      Suv320×2400.0430.051(93,42)(87,38)
    • Table 3. Accuracy comparison of proposed algorithm and DiMP

      View table
      View in Article

      Table 3. Accuracy comparison of proposed algorithm and DiMP

      序列名称序列总帧数IoU准确率
      本文DiMP
      BlurCar17410.50.9970.991
      0.70.9420.906
      BlurCar25840.51.0000.995
      0.70.9950.945
      BlurCar33560.51.0000.985
      0.71.0000.966
      BluCar43790.51.0001.000
      0.70.9820.982
      Car110190.50.9960.985
      0.70.7620.645
      Car29120.51.0001.000
      0.70.9700.970
      Car46580.51.0001.000
      0.70.9790.958
      Car2430580.50.9990.999
      0.70.9500.930
      CarDark3920.50.9950.959
      0.70.5700.418
      CarScale2510.51.0001.000
      0.70.9880.972
      SUV9440.50.9440.907
      0.70.7660.748
      平均结果0.50.9930.983
      0.70.9000.858
    • Table 4. Frame rate comparison of proposed algorithm and DiMP

      View table
      View in Article

      Table 4. Frame rate comparison of proposed algorithm and DiMP

      算法平台帧率/f/s帧率功耗比/f/s/W权重文件大小/MB平均准确率
      IoU(0.5)IoU(0.7)
      DiMPTiTan X(250 W)400.163630.9830.858
      本文算法Atlas200(12 W)453.332780.9930.900
      本文算法(通道剪枝)Atlas200(12 W)665.55900.9910.885
    • Table 5. Frame rates on different platforms

      View table
      View in Article

      Table 5. Frame rates on different platforms

      平台帧率/f/s帧率/f/s(通道剪枝)
      NVIDIA TX21221
      NVIDIA Xaiver2235
      Atlas2004566
    Tools

    Get Citation

    Copy Citation Text

    Xi HE, Liangfu LI, Jiaoying WANG, Jie WANG, Xiaoyan LU, Jun QIAN, Yizhou YANG, Peizhen LIU. Implementation of real-time deep learning tracking algorithm based on NPU[J]. Journal of Applied Optics, 2022, 43(4): 682

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: OE INFORMATION ACQUISITION AND PROCESSING

    Received: Feb. 21, 2022

    Accepted: --

    Published Online: Aug. 10, 2022

    The Author Email:

    DOI:10.5768/JAO202243.0402003

    Topics