Please enable JavaScript.

Coggle requires JavaScript to display documents.

Object Detection by using Automotive Radar(One Step), Transformer Encoder…

- - - - Range Doppler Map based
        
        One Stage Object Detector based(Not Feasible for Real Application Since Different Objects in Azimuth Can Not be Recognized).
      - Range-Azimuth Heatmap based
        
        One Stage Object Detector based
      - Multi-View(Range-Azimuth, Range-Doppler, Dopper-Azimuth) based
        
        Two Stage Object Detector based, One for Range Detection and One for Angle Detection(Not Same as Two Stage Object Detector in Computer Vision)
        
        One Stage Object Detector based
    - - Processing Radar Detection Points Directly(With Accumulation Over Multi-Radar)
        
        Modified PointNet/PointNet++ based Object Detector, Use Each Point as Anchor Point(Either Anchor Point Only or Several Anchor Boxes Affilicated to Each Anchor Point) to Estimate Confidence Score of Having Object for Each Anchor Point and Regression of Bounding Boxes based on Every Anchor Point.
      - using Pseudo Image Discretized from Radar Detection Points
        
        One Stage Object Detector based
  - - - Range-Azimuth Heatmap in Time based
        
        using 3D Conv CNN based Encoder Decoder Framework One Stage Object Detector to Process 3D Data(Range, Azimuth, Time).
        
        Concatenate Extracted Features from Different Frames(using Encoder Decoder Framework) Together, Then Decoding and Object Detection.
        
        Time Series Processing of Extracted Features from Different Frames(using Encoder Decoder Framework), e.g. Non-Markovian Chain Model based LSTM, Then Object Detection.
      - Multi-View(Range-Azimuth, Range-Doppler, Dopper-Azimuth) in Time based
    - - Processing Radar Detection Points Directly(With Accumulation Over Time)
        
        Similar Way of Processing as Single-Frame (Multi-Radar) Radar Detection Points based Solution(but for Multi-Frame Radar Detection Points Solution, Linear Interpolation of Radar Detection Points from Previous Frames to Current Frame is Needed)
      - using Pseudo Image Discretized from Radar Detection Points, Similar Way of Processing as Multi-Frame Low Level Radar Data based Solutions(Just Input "Image" is Different)