This solution duplicates the camera source data to two TDA4x SoCs, (TDA4-A, TDA4-B) each TDA4x processor implements different ADAS functions. A reference implementation to distribute front camera functionality, as shown in Figure 3-1.

Camera in Primary TDA4-A:

• Front camera is used to monitor the front camera object:
  • The camera data input through CSI2-1 port
  • Preprocessing by C66x DSP (such as Format/Grayscale conversion, ROI Setting)
  • Run inference on deep learning models to achieve Lane Departure Warning (LDW) and Lane Keeping Assist (LKA) function in C7x+MMA deep learning accelerator
  • Postprocessing by C66x DSP (such as drawing line and interpolation) and output the result to fusion node
• Side view cameras are used to monitor car and objects coming from the side:
  • The camera data input through CSI2-0 port
  • Preprocessing/postprocessing by C66x DSP
  • Run inference on deep learning models for Lane Changing Assist (LCA) and Adaptive Cruise Control (ACC) in C7x+MMA
  • Output the result to fusion node

Camera in Secondary TDA4-B:

• Front camera is used to monitor the front object:
  • The camera data input through CSI2-0 port
  • Preprocessing/postprocessing is done by C66x DSP
  • Run inference deep learning models to achieve Object Detect (OD) function in C7x+MMA, such as Traffic Light Recognition (TLR), Traffic Sign Recognition (TSR)
  • Output the result to PCIe TX node
• Surround view cameras are used to detect the environment around the car at low speed:
  • The camera data input through CSI2-1 port
  • The codec node is optional for recording (DVR)
  • GPU node for 3D rendering
  • Preprocessing/postprocessing by C66x DSP
  • Run inference deep learning models to achieve Automated Valet Parking (AVP) and Automated Parking Assist (APA) function in C7x+MMA
  • Output the result to PCIe TX node

Fusion:

• All of the results from the TDA4-B are transferred to the TDA4-A via PCIe TX/RX node, then input to fusion node.
• All of the results whether from TDA4-A or TDA4-B are assembled in fusion node, then output the finally result.