International Core Journal of Engineering 2020-26 | Page 91

TABLE I.
Identifier
Right
Left
Straight right turn
D ATABASE COMPOSITION
sample graph
Quantity
9771
9733
7718
Straight left turn 7636
Straight 14645
Crosswalk 12647
The images of the data set were randomly scrambled
before training, and the expanded data was divided into
training set, verification set and test set according to the ratio
of 6:2:2, and the number was 22087, 7362, 7362 respectively.
conclusions can be reflected. Here we named the test set that
was randomly divided proportionally as test set 1, and the test
set taken on Huashan Road was named test set 2.
3) Experimental results and analysis
SSD, SSD (improved loss function), FSSD, and R-FCN
models are all based on the Python interface under the Caffe
framework. The operating system is Ubuntu 16.04 and the
GPU is Nvidia 1080Ti. Test the two test sets separately under
the premise of model convergence. The test results are shown
in Table II and Table III.
This paper prepares another test set that captures 848
images taken on a car camera on Huashan Road. Different
from the previous data set, the test set randomly selects the
road information of a specific location in China, which is more
similar to the actual detection scenario. If the results of the two
test sets are similar, the general applicability of the
TABLE II. RESULTS OF DIFFERENT MODELS IN SET 1
model
SSD SSD(Focal loss) FSSD R-FCN
0.619
0.623
0.620
0.631
0.859
0.897
0.708
33 0.627
0.659
0.638
0.620
0.860
0.906
0.718
33 0.668
0.657
0.669
0.659
0.887
0.916
0.743
41 0.911
0.905
0.898
0.899
0.906
0.924
0.907
56
category
Right
Left
Straight right
Straight left
Straight
Crosswalk
mAP
speed (ms)
T ABLE III RESULTS OF DIFFERENT MODELS IN SET 1
model
SSD SSD(Focal loss) FSSD R-FCN
0.619
0.623
0.620
0.631
0.859
0.897
0.708
33 0.627
0.659
0.638
0.620
0.860
0.906
0.718
33 0.668
0.657
0.669
0.659
0.887
0.916
0.743
41 0.911
0.905
0.898
0.899
0.906
0.924
0.907
56
category
Right
Left
Straight right
Straight left
Straight
Crosswalk
mAP
speed (ms)
Through the above results, it can be noticed that the single-
stage detection algorithm such as R-FCN has higher overall
accuracy, and the two-stage detection algorithm such as SSD
has lower overall accuracy. According to the average
detection accuracy, the order from high to low is R-FCN,
FSSD, SSD (Focal loss), SSD. SSD, SSD (Focal loss), FSSD
algorithm is poor in detecting arrow marks. In terms of speed,
the single-stage algorithm is faster and can meet real-time
performance. The R-FCN is slower, but the gap is not large.
and one-stage algorithms such as SSD do not utilize shallow
features. The SSD/FSSD input fixed size of 300*300
increases the speed, but has a large impact on the small target
detection of large images. For large target detection RFCN
and ssd/fssd performance is equivalent. Using two different
data sets for model testing, it can be seen from the test results
that the accuracy is basically the same, which shows the
accuracy and universal adaptability of the test results. In
general, the R-FCN model is the best for road marking
detection.
V. C ONCLUSION
R EFERENCES
For small targets, detecting R-FCN is better. For large
targets such as crosswalks, the two accuracy rates are
comparable, but the onestage method is faster. Small target
detection tends to rely more on shallow features since shallow
features have high resolution but poor semantic distinction,
[1]
[2]
69
H. Vokhidov, H. Hong, K. Jin, , et al, “Recognition of Damaged
Arrow-Road Markings by Visible Light Camera Sensor Based on
Convolutional Neural Network”, Sensors, vol. 16, no. 12, 2016.
R. Girshick, J. Donahue, T. Darrell, J. Malik, “Rich feature hierarchies