Skip to content

Latest commit

 

History

History
83 lines (60 loc) · 2.1 KB

README.md

File metadata and controls

83 lines (60 loc) · 2.1 KB

Lane_Detection_on_KylinPhytium-2000/4

This project is based on the KylinPhytium-2000/4 platform and used to detect lanes in videos or cameras. The project includes the following parts:
--image preprocessing based on algorithm from [LIME: Low-Light Image Enhancement via Illumination Map Estimatio] (https://ieeexplore.ieee.org/document/7782813);
--lane detection based on network models UNet and Transformer;
--GUI developed with QT.

Several achievements we have made:
--Reconstruct the LIME preprocessing algorithm with C++;
--Use NEON instructions and OpenMP API to speed up the LIME algorithm;
--Rewrite the Fouriour transform algorithm instead of cv::dft and cv::idft to improve the performance of the LIME algorithm;
--Use Depthwise Separable Convolution and Quantization on UNet to reduce the model size and speed up the inference;
--Deploy UNet on the NCNN framework and transformer on the onnx framework to speed up the inference;
--Develop a realtime monitor module in GUI to display the hardware performance including CPU and memory usage.

Requirements

  • KylinPhytium-2000/4
  • OpenCV 4.7.0
  • NCNN
  • onnxruntime
  • Qt 5.12.8

Usage

1.LIME preprocessing

  • mkdir build
  • cd build
  • cmake ..
  • make
  • ./lime

2.Network inference

  • mkdir build
  • cd build
  • cmake ..
  • make
  • ./unet_ncnn ../images/0.jpg for UNet
  • ./LSTR for Transformer

3.GUI Need QT creator based on Qt 5.12.8

Performance

For the LIME preprocessing, the performance is shown as follows:

Before LIME preprocessing:

After LIME preprocessing:

For the network inference, the performance is shown as follows:

UNet:

Transformer:

The GUI is shown as follows:

Demo

The demo is shown in the video:

demo.mp4