This repo is for deploy Ultra Fast Lane Detection in C++.
Hardware: Nvidia AGX Xavier(ARM)
Software:
- JetPack 4.4
- CUDA 10.2
- CUDNN 8.0.0
- TensorRT 7.1.3
Tested with above but can be transfered to other environments.
- Build:
mkdir build && cd build && cmake .. && make
- Model conversion:
- Download lane.wts from model (code:14je) and put it into models/
- Convert to tensorrt engine
./model_conversion
- Lane_detection:
./main
- Input: Define variable
video
inUFLD.h
to be the input video file path - Output: In this algorithm, row locations of lanes are predefined as
tusimple_row_anchor
inUFLD.h
, column locations are put intolanes
which is a 2-dimensional array with shape (4,56), 4 is the number of lanes, 56 is the number of row-anchors, to get one lane represented by 56 points, refer to function 'display()' inUFLD.cpp
.