Lane Detection Optimized with GPU Coder

This example shows how to generate CUDA® code from a deep learning network, represented by a SeriesNetwork object. The SeriesNetwork in this example is a convolutional neural network that can detect and output lane marker boundaries from image.


  • CUDA enabled NVIDIA® GPU with compute capability 3.2 or higher.

  • NVIDIA CUDA toolkit and driver.

  • NVIDIA cuDNN library v7 or higher.

  • OpenCV 3.1.0 libraries for video read and image display operations.

  • Environment variables for the compilers and libraries. For information on the supported versions of the compilers and libraries, see Third-party Products. For setting up the environment variables, see Environment Variables.

  • Deep Learning Toolbox™ for using SeriesNetwork objects.

  • GPU Coder™ for generating CUDA code.

  • GPU Coder Interface for Deep Learning Libraries support package. To install this support package, use the Add-On Explorer.

Verify the GPU Environment

Use the coder.checkGpuInstall function and verify that the compilers and libraries needed for running this example are set up correctly.


Get the Pretrained SeriesNetwork

[laneNet, coeffMeans, coeffStds] = getLaneDetectionNetwork();

This network takes image as an input and outputs two lane boundaries that correspond to the left and right lanes of the ego vehicle. Each lane boundary can be represented by the parabolic equation: , where y is the lateral offset and x is the longitudinal distance from the vehicle. The network outputs the three parameters a,b, and c per lane. The network architecture is similar to AlexNet, except that the last few layers are replaced by a smaller fully connected layer and regression output layer.

ans = 

  23x1 Layer array with layers:

     1   'data'          Image Input                   227x227x3 images with 'zerocenter' normalization
     2   'conv1'         Convolution                   96 11x11x3 convolutions with stride [4  4] and padding [0  0  0  0]
     3   'relu1'         ReLU                          ReLU
     4   'norm1'         Cross Channel Normalization   cross channel normalization with 5 channels per element
     5   'pool1'         Max Pooling                   3x3 max pooling with stride [2  2] and padding [0  0  0  0]
     6   'conv2'         Convolution                   256 5x5x48 convolutions with stride [1  1] and padding [2  2  2  2]
     7   'relu2'         ReLU                          ReLU
     8   'norm2'         Cross Channel Normalization   cross channel normalization with 5 channels per element
     9   'pool2'         Max Pooling                   3x3 max pooling with stride [2  2] and padding [0  0  0  0]
    10   'conv3'         Convolution                   384 3x3x256 convolutions with stride [1  1] and padding [1  1  1  1]
    11   'relu3'         ReLU                          ReLU
    12   'conv4'         Convolution                   384 3x3x192 convolutions with stride [1  1] and padding [1  1  1  1]
    13   'relu4'         ReLU                          ReLU
    14   'conv5'         Convolution                   256 3x3x192 convolutions with stride [1  1] and padding [1  1  1  1]
    15   'relu5'         ReLU                          ReLU
    16   'pool5'         Max Pooling                   3x3 max pooling with stride [2  2] and padding [0  0  0  0]
    17   'fc6'           Fully Connected               4096 fully connected layer
    18   'relu6'         ReLU                          ReLU
    19   'drop6'         Dropout                       50% dropout
    20   'fcLane1'       Fully Connected               16 fully connected layer
    21   'fcLane1Relu'   ReLU                          ReLU
    22   'fcLane2'       Fully Connected               6 fully connected layer
    23   'output'        Regression Output             mean-squared-error with 'leftLane_a', 'leftLane_b', and 4 other responses

Examine Main Entry Point Function

type detect_lane.m
function [laneFound, ltPts, rtPts] = detect_lane(frame, laneCoeffMeans, laneCoeffStds) 
% From the networks output, 
% compute left and right lane points in the image coordinates 
% The camera coordinates are described by the caltech mono camera model.


persistent lanenet;
if isempty(lanenet)
    lanenet = coder.loadDeepLearningNetwork('laneNet.mat', 'lanenet');

lanecoeffsNetworkOutput = lanenet.predict(permute(frame, [2 1 3]));

% Recover original coeffs by reversing the normalization steps

params = lanecoeffsNetworkOutput .* laneCoeffStds + laneCoeffMeans;

isRightLaneFound = abs(params(6)) > 0.5; %c should be more than 0.5 for it to be a right lane
isLeftLaneFound =  abs(params(3)) > 0.5;

vehicleXPoints = 3:30; %meters, ahead of the sensor
ltPts = coder.nullcopy(zeros(28,2,'single'));
rtPts = coder.nullcopy(zeros(28,2,'single'));

if isRightLaneFound && isLeftLaneFound
    rtBoundary = params(4:6);		
	rt_y = computeBoundaryModel(rtBoundary, vehicleXPoints);
	ltBoundary = params(1:3);
	lt_y = computeBoundaryModel(ltBoundary, vehicleXPoints);
	% Visualize lane boundaries of the ego vehicle
    tform = get_tformToImage;
    % map vehicle to image coordinates
    ltPts =  tform.transformPointsInverse([vehicleXPoints', lt_y']);
    rtPts =  tform.transformPointsInverse([vehicleXPoints', rt_y']);
	laneFound = true;
	laneFound = false;


function yWorld = computeBoundaryModel(model, xWorld)
	yWorld = polyval(model, xWorld);	

function tform = get_tformToImage 
% Compute extrinsics based on camera setup
yaw = 0;
pitch = 14; % pitch of the camera in degrees
roll = 0;

translation = translationVector(yaw, pitch, roll);
rotation    = rotationMatrix(yaw, pitch, roll);

% Construct a camera matrix
focalLength    = [309.4362, 344.2161];
principalPoint = [318.9034, 257.5352];
Skew = 0;

camMatrix = [rotation; translation] * intrinsicMatrix(focalLength, ...
	Skew, principalPoint);

% Turn camMatrix into 2-D homography
tform2D = [camMatrix(1,:); camMatrix(2,:); camMatrix(4,:)]; % drop Z

tform = projective2d(tform2D);
tform = tform.invert();

function translation = translationVector(yaw, pitch, roll)
SensorLocation = [0 0];
Height = 2.1798;    % mounting height in meters from the ground
rotationMatrix = (...
	rotZ(yaw)*... % last rotation
	rotZ(roll)... % first rotation

% Adjust for the SensorLocation by adding a translation
sl = SensorLocation;

translationInWorldUnits = [sl(2), sl(1), Height];
translation = translationInWorldUnits*rotationMatrix;

% Rotation around X-axis
function R = rotX(a)
a = deg2rad(a);
R = [...
	1   0        0;
	0   cos(a)  -sin(a);
	0   sin(a)   cos(a)];


% Rotation around Y-axis
function R = rotY(a)
a = deg2rad(a);
R = [...
	cos(a)  0 sin(a);
	0       1 0;
	-sin(a) 0 cos(a)];


% Rotation around Z-axis
function R = rotZ(a)
a = deg2rad(a);
R = [...
	cos(a) -sin(a) 0;
	sin(a)  cos(a) 0;
	0       0      1];

% Given the Yaw, Pitch and Roll, figure out appropriate Euler
% angles and the sequence in which they are applied in order to
% align the camera's coordinate system with the vehicle coordinate
% system. The resulting matrix is a Rotation matrix that together
% with Translation vector define the camera extrinsic parameters.
function rotation = rotationMatrix(yaw, pitch, roll)

rotation = (...
	rotY(180)*...            % last rotation: point Z up
	rotZ(-90)*...            % X-Y swap
	rotZ(yaw)*...       % point the camera forward
	rotX(90-pitch)*...  % "un-pitch"
	rotZ(roll)...       % 1st rotation: "un-roll"

function intrinsicMat = intrinsicMatrix(FocalLength, Skew, PrincipalPoint)
intrinsicMat = ...
	[FocalLength(1)  , 0                     , 0; ...
	 Skew             , FocalLength(2)   , 0; ...
	 PrincipalPoint(1), PrincipalPoint(2), 1];

Generate Code for Network and Post Processing Code

The network computes parameters a, b, and c describing the parabolic equation for the left and right lane boundaries.

From these parameters, we can compute the x and y coordinates corresponding to the lane positions. Furthermore, the coordinates must be mapped to image coordinates. The function detect_lane.m does all these computations. We can generate CUDA code for this function as well.

cfg = coder.gpuConfig('lib');
cfg.DeepLearningConfig = coder.DeepLearningConfig('cudnn');
cfg.GenerateReport = true;
cfg.TargetLang = 'C++';
codegen -args {ones(227,227,3,'single'),ones(1,6,'double'),ones(1,6,'double')} -config cfg detect_lane
Code generation successful: To view the report, open('codegen/lib/detect_lane/html/report.mldatx').

Generated Code Description

The SeriesNetwork is generated as a C++ class containing an array of 23 layer classes.

class c_lanenet
  int32_T batchSize;
  int32_T numLayers;
  real32_T *inputData;
  real32_T *outputData;
  MWCNNLayer *layers[23];
  void setup(void);
  void predict(void);
  void cleanup(void);

The setup() method of the class sets up handles and allocates memory for each layer object. The predict() method invokes prediction for each of the 23 layers in the network.

The files cnn_lanenet_conv*_w and cnn_lanenet_conv*_b are the binary weights and bias file for convolution layer in the network. The files cnn_lanenet_fc*_w and cnn_lanenet_fc*_b are the binary weights and bias file for fully connected layer in the network.

codegendir = fullfile('codegen', 'lib', 'detect_lane');
.                              cnn_lanenet_conv5_b            
..                             cnn_lanenet_conv5_w            cnn_lanenet_fc6_b              
DeepLearningNetwork.h          cnn_lanenet_fc6_w              
DeepLearningNetwork.o          cnn_lanenet_fcLane1_b               cnn_lanenet_fcLane1_w          
MWCNNLayerImpl.hpp             cnn_lanenet_fcLane2_b          
MWCNNLayerImpl.o               cnn_lanenet_fcLane2_w             cnn_lanenet_responseNames.txt  
MWCudaDimUtility.h             codeInfo.mat                   
MWCudaDimUtility.o             detect_lane.a                  
MWFusedConvReLULayer.hpp       detect_lane.h                  
MWFusedConvReLULayer.o         detect_lane.o               
MWFusedConvReLULayerImpl.hpp   detect_lane_initialize.h       
MWFusedConvReLULayerImpl.o     detect_lane_initialize.o         detect_lane_ref.rsp            
buildInfo.mat                  detect_lane_terminate.h        
cnn_api.cpp                    detect_lane_terminate.o        
cnn_api.hpp                    detect_lane_types.h            
cnn_api.o                      examples                       
cnn_lanenet_avg                gpu_codegen_info.mat           
cnn_lanenet_conv1_b            html                           
cnn_lanenet_conv1_w            interface                      
cnn_lanenet_conv2_w            predict.h                      
cnn_lanenet_conv3_b            predict.o                      
cnn_lanenet_conv3_w            rtw_proj.tmw                   
cnn_lanenet_conv4_b            rtwtypes.h                     

Generate Additional Files Needed for Post Processing Output

% export mean and std values from the trained network for use in testing
codegendir = fullfile(pwd, 'codegen', 'lib','detect_lane');
fid = fopen(fullfile(codegendir,'mean.bin'), 'w');
A = [coeffMeans coeffStds];
fwrite(fid, A, 'double');

Main File

The network code is to be compiled with a main file.

The main file uses the OpenCV VideoCapture method to read frames from the input video. Each frame is processed and classified, until no more frames are to be read. Before displaying the output for each frame, the outputs are post-processed using the detect_lane function generated in detect_lane.cpp.

type main_lanenet.cpp
/* Copyright 2016 The MathWorks, Inc. */

#include <stdio.h>
#include <stdlib.h>
#include <cuda.h>
#include "opencv2/opencv.hpp"
#include <list>
#include <cmath>
#include "detect_lane.h"

using namespace cv;
void readData(float *input, Mat& orig, Mat & im)
	Size size(227,227);
	for(int j=0;j<227*227;j++)
		//BGR to RGB

void addLane(float pts[28][2], Mat & im, int numPts)
    std::vector<Point2f> iArray;
    for(int k=0; k<numPts; k++) 
    Mat curve(iArray, true);
    curve.convertTo(curve, CV_32S); //adapt type for polylines
    polylines(im, curve, false, CV_RGB(255,255,0), 2, CV_AA);

void writeData(float *outputBuffer, Mat & im, int N, double means[6], double stds[6])
    // get lane coordinates
    boolean_T laneFound = 0;	
    float ltPts[56];
    float rtPts[56];	
    detect_lane(outputBuffer, means, stds, &laneFound, ltPts, rtPts);    
	if (!laneFound)
	float ltPtsM[28][2];
	float rtPtsM[28][2];
	for(int k=0; k<28; k++)
		ltPtsM[k][0] = ltPts[k];
		ltPtsM[k][1] = ltPts[k+28];
		rtPtsM[k][0] = rtPts[k];
		rtPtsM[k][1] = rtPts[k+28];   

	addLane(ltPtsM, im, 28);
	addLane(rtPtsM, im, 28);

void readMeanAndStds(const char* filename, double means[6], double stds[6])
    FILE* pFile = fopen(filename, "rb");
    if (pFile==NULL)
        fputs ("File error",stderr);

    // obtain file size
    fseek (pFile , 0 , SEEK_END);
    long lSize = ftell(pFile);
    double* buffer = (double*)malloc(lSize);
    size_t result = fread(buffer,sizeof(double),lSize,pFile);
    if (result*sizeof(double) != lSize) {    
        fputs ("Reading error",stderr);
    for (int k = 0 ; k < 6; k++)
        means[k] = buffer[k];
        stds[k] = buffer[k+6];

// Main function
int main(int argc, char* argv[])
    float *inputBuffer = (float*)calloc(sizeof(float),227*227*3);
    float *outputBuffer = (float*)calloc(sizeof(float),6);

    if ((inputBuffer == NULL) || (outputBuffer == NULL)) {
        printf("ERROR: Input/Output buffers could not be allocated!\n");
    // get ground truth mean and std
    double means[6];
    double stds[6];	
    readMeanAndStds("mean.bin", means, stds);	
	if (argc < 2)
        printf("Pass in input video file name as argument\n");
        return -1;
    VideoCapture cap(argv[1]);
    if (!cap.isOpened()) {
        printf("Could not open the video capture device.\n");
        return -1;

    cudaEvent_t start, stop;
    float fps = 0;
    Mat orig, im;    
    namedWindow("Lane detection demo",CV_WINDOW_NORMAL);
        cap >> orig;
        if (orig.empty()) break;                
        readData(inputBuffer, orig, im);		

        writeData(inputBuffer, orig, 6, means, stds);
        char strbuf[50];
        float milliseconds = -1.0; 
        cudaEventElapsedTime(&milliseconds, start, stop);
        fps = fps*.9+1000.0/milliseconds*.1;
        sprintf (strbuf, "%.2f FPS", fps);
        putText(orig, strbuf, cvPoint(200,30), CV_FONT_HERSHEY_DUPLEX, 1, CV_RGB(0,0,0), 2);
        imshow("Lane detection demo", orig); 		
        if( waitKey(50)%256 == 27 ) break; // stop capturing by pressing ESC	*/       
    destroyWindow("Lane detection demo");
    return 0;

Download the Example Video

if ~exist('./caltech_cordova1.avi', 'file')
	url = '';
	websave('caltech_cordova1.avi', url);

Build the Executable

if ispc
    setenv('MATLAB_ROOT', matlabroot);
    vspath = mex.getCompilerConfigurations('C++').Location;
    setenv('VSPATH', vspath);
    system('lanenet.exe ..\..\..\caltech_cordova1.avi');
    setenv('MATLAB_ROOT', matlabroot);
    system('make -f');
    system('./lanenet ../../../caltech_cordova1.avi');

Input Screenshot

Output Screenshot