TI Autonomous Driving Algorithms (TIADALG) Library User Guide
TIADALG: Motion Segmentation and Object Detection Design

Requirements Addressed {#did_TIADALG motion segmentation and object detection requirements}

Introduction

Purpose

Short Application Description

The models include semantic segmentation, motion segmentation and a joint model that does both together. The models are trained in PyTorch and were finally converted to ONNX format. Cityscapes dataset was used for the training and the accuracy numbers reported are on this dataset.

Input and Output format

Pre-Processing for input Image

  • Step 1: Resize to 1024x512
  • Step 2: Divide by 255.0
  • Step 3: Mean subtract ([0.485, 0.456, 0.406])
  • Step 4: Multiply ([1/0.229, 1/0.224, 1/0.225])

Semantic Segmentation

  • Input: RGB Image preprocessed as dscussed above
  • Output: integer output plane, indicating the class id of segmentation.
  • Validation Accuracy: 70.4 mIoU

Motion Segementation

  • Input1: RGB Image preprocessed as dscussed above
  • Input2: 3 planes of DOF data
  • Output : integer output plane, indicating the class id of segmentation.
  • Validation Accuracy: 85.0 mIoU

Directory Structure

Diagrams

Sequence Diagram

Component Interaction

OpenVX Graph

Resource usage

Error handling

Interface

Design Analysis and Resolution (DAR)

Design Decision : none

na

Design Criteria: none

na

Design Alternative: none

na

Design Alternative: none

na

Final Decision

na


© Copyright 2018 Texas Instruments Incorporated. All rights reserved.
Document generated by doxygen 1.8.6