SPRADH2A February   2024  – November 2024 AM62A3 , AM62A3-Q1 , AM62A7 , AM62A7-Q1 , AM62P , AM62P-Q1 , DS90UB953A-Q1 , DS90UB960-Q1 , TDES960 , TSER953

 

  1.   1
  2.   Abstract
  3.   Trademarks
  4. 1Introduction
  5. 2Connecting Multiple CSI-2 Cameras to the SoC
    1. 2.1 CSI-2 Aggregator Using SerDes
    2. 2.2 CSI-2 Aggregator without Using SerDes
    3. 2.3 Supported Camera Data Throughput
  6. 3Enabling Multiple Cameras in Software
    1. 3.1 Camera Subsystem Software Architecture
    2. 3.2 Image Pipeline Software Architecture
  7. 4Reference Design
    1. 4.1 Supported Cameras
    2. 4.2 Setting up Four IMX219 Cameras
    3. 4.3 Configuring Cameras and CSI-2 RX Interface
    4. 4.4 Streaming from Four Cameras
      1. 4.4.1 Streaming Camera Data to Display
      2. 4.4.2 Streaming Camera Data through Ethernet
      3. 4.4.3 Storing Camera Data to Files
    5. 4.5 Multicamera Deep Learning Inference
      1. 4.5.1 Model Selection
      2. 4.5.2 Pipeline Setup
  8. 5Performance Analysis
  9. 6Summary
  10. 7References
  11. 8Revision History

Model Selection

TI’s EdgeAI-ModelZoo provides hundreds of state-of-the-art models which are converted or exported from the original training frameworks to an embedded friendly format so that the models can be offloaded to the C7x-MMA deep learning accelerator. The cloud-based Edge AI Studio provides an easy-to-use Model Selection tool. This is dynamically updated to include all models supported in TI EdgeAI-ModelZoo. The tool requires no previous experience and provides an easy-to-use interface to enter the features required in the desired model.

The TFL-OD-2000-ssd-mobV1-coco-mlperf was selected for this multi-camera deep learning experiment. This multi-object detection model is developed in the Tensor Flow framework with 300x300 input resolution. Table 4-1 shows the important features of this model when trained on the coco dataset with about 80 different classes.

Table 4-1 Highlight Features of the Model TFL-OD-2000-ssd-mobV1-coco-mlperf.
ModelTaskResolutionFPSmAP 50% Accuracy On COCOLatency/Frame (ms)DDR BW Utilization (MB/Frame)
TFL-OD-2000-ssd-mobV1-coco-mlperf Multi Object Detection300x300~15215.96.518.839