Analysis of [paper] pointlanenet papers

Catalog

Preface

abstract

method description

Structured representation of lane lines

Distance

detection pipeline

Model Analysis

Training

experiments

  summary

---

Preface

PointLaneNet yes 2019 IEEE Intelligent Vehicles Symposium (IV) A paper on , Determine the location of the lane line by detecting the points on the lane line , And classify the lane lines , Is subject to YOLO Series inspired ont-stage Network structure .

Title:《PointLaneNet: Efficient end-to-end CNNs for Accurate Real-Time Lane Detection》

paper

version 1.0 

abstract

An end-to-end lane detection algorithm is proposed , The coordinates of lane line points can be obtained directly from the driving scene .

method description

Inspired by YOLO Series algorithm , The lane detection algorithm can simultaneously predict the location and classify the lane lines in a single network , In order to facilitate network training , The concept of lane line distance is also clarified .

Structured representation of lane lines

The article points out that ： Generally, straight lines or polynomial curves are used to structurally represent lane lines , In fact, in reality , This is not advisable , Even the curve of cubic polynomial cannot completely represent the complex lane line , But a series of precise points can represent the lane line , These points can be obtained by interpolation , As long as you know the starting point in the lane line , Termination point , In the middle , The four parameters of lane line width , It can show the shape of the lane line . The width of the lane line is known （ According to the regulations ）, The end point of the lane line is at the bottom of the picture （ This article defaults to ）.

There is a unique intersection between the horizontal and equally spaced straight lines and each lane line , Now you only need to find the starting point of the lane line and the point of the lane line .Fig2 Is the representation of structured lane lines .

​

 y It's pre-set （ Horizontal line ）, So if you want to determine the coordinates of each point, you only need to determine the offset of that point , therefore , The predicted value of each lane line is ： Offset , The starting point , Confidence of lane line . The following figure is the definition of points on the lane line ：

​

Distance

Define the distance between lane lines （ The paper says ： The definition of the distance between lane lines is very important for training . In a general way ,point-point Distance is Euclidean distance ,box-box yes IOU Definition , The distance between lane lines is not clearly defined ）.

Yes point-line Distance is defined ：

Euclidean distance .

​

Yes  line-line Distance definition formula ：

The distance between two lines is defined as the average of the distance between overlapping points . For non overlapping parts , The distance defined is infinite （ infinity ）.

​

 detection pipeline

Use one-stage Methods to detect the lane line , The model detects the location of points and classifies lane lines .

​

Use GoogleNet As feature extraction network , Conduct 8 Next sampling , The pictures are divided into w x h Of grid  

Location detection

Use 2 Time 1 x1 Convolution of , take feature map The output dimension of is converted to n+1,n yes n Offset , another 1 Dimension is the starting position of the lane line .

​

Classification

Use 2 Time 1 x1 Convolution of , Convert the output channel to 2, To indicate whether the lane line passes through this grid, By modifying the channel , This can be used to detect the color or type of lane line , Because of the dataset used in this article , Here is just to detect whether the lane line exists or not . The final output is ：w * h * (n+1+2)

post-processing:

Because a lane passes through many grid, Each lane line has many prediction results , So it needs post-processing .

Specific steps ：

1. Initialization distance , The threshold for \delta = 10 pixels,confidence The threshold of is cls = 0.99;

2. Delete all cls < 0.99 The lane line of the road ;

3. Yes confidence Sort , selection confidence The tallest one , If confidence The highest one distance < \delta, Delete this lane line , Keep looking for the second highest confidence;

4. If lane number More than the 5, take \delta = \delta+10, Re step 3; Or it's over ;

notes ： This data set is set with no more than five lanes ;

Be careful ： Formula is Latex,\delta.

Model Analysis

Model evaluation index :

Yes Mem、 Amount of computation FLOPS And the computational strength of the model I Quantitative analysis . among , The computational strength of the model shows Floating point operands required for memory exchange per byte during calculation , Unit is Flops/Byte. The relationship between the three variables ：

I = Flops/Mem

model Parameter comparison

contrast SCNN,LaneNet.( The article here uses three lightweight feature networks as a comparison ,GoogLeNet, SqueezeNet, and MobileNet)

Training

data: The operation of data enhancement ;

Matching strategy ： To balance positive and negative samples , In training set , Set the grid center point within a certain distance threshold from the true lane line as a positive sample ; Otherwise, set as negative sample .threshold = 20 pixels.

Loss： Three kinds of predicted values output by the model , Confidence of lane line confidence, Offset of lane line point \triangle x（x-coordinate）, Starting point of lane line y（y-coordinate）, Correspondingly ,loss Function is also composed of these three parts , Of confidence loss Using cross entropy , The loss function of lane line points uses Euclidean distance .loss function As shown in the figure below ：

Choosing scales： Determine the sampling magnification through experiments , This paper chooses 16 Double down sampling , Although the information of lane line may be lost after down sampling , However, the location of the lane line can still be known through the lane structure or environmental information .

Hard negative mining： Sort the negative samples according to the confidence , Select the sample with high score , And adjust the ratio of positive samples to negative samples to 1:3, To strengthen the network's ability to identify false positives .（ I don't understand here , Leave a hole ）

experiments

Used Tusimple Data set and CuLane Data sets .

The evaluation index ：accuracy,FP,FN

 

  summary

PointLaneNet yes one-stage Network structure , A network outputs the location information of lane lines and lane line classification information at the same time , Create a horizontal line on the input image , There is a unique intersection between the straight line and the lane line , By finding the starting point of the lane line , Termination point , The center point and the width of the lane line determine the shape of the lane line and the classification of the lane , The final output of the network is ： Offset , The starting point , Confidence of lane line .（ y It's pre-set （ Horizontal line ）, So if you want to determine the coordinates of each point, you only need to determine the offset of that point ）; Besides , Defines the lane line and between the lane lines , Distance between point and lane line ; Set a threshold to determine whether the detected lane line is a positive sample or a negative sample .