Page 33 - สำนักบำรุงทาง รายงานประจำปี 2563

(a) Vid 1 (noisy video) (b) Vid 2 (high light intensity at 3PM)

(c) Vid 3 (proper light intensity at 5PM) (d) Vid 4 (night time at 3AM)

Figure 2. Four experimental sample image frames of the four video streams
used for the validation of detection and count.truck and trailer respectively.

Table 1. Validation of count in terms of quality of video
Video Manual Method
Video OA
duration Count Count
Figure 1. Sample results produced by the trained YOLOv5l model on test images. Vid 1 (noisy video) 1 hour 3617 3390 93.7
The classes are predicted from 0 to 6 for car, bus, taxi, bike, pickup, truck and trailer respectively. Vid 2 (high light intensity at 3PM) 5 min 552 545 98.7
To implement the method and carry-out our experiments, we used a setup of several cameras in the Vid 3 (proper light intensity at 5PM) 5 min 514 512 99.6
highway of Ratchapruek, Pathum Thani, Thailand. We divide the validation of our method into three parts as Vid 4 (night time at 3AM) 5 min 14 12 85.7
validation of count, classification and speed. As the performance of the method could possibly be hampered by Total Accuracy 94.4
the quality of streamed video, we clip four videos from the real-time stream from four cameras of different image
quality for the purpose of validation of count as shown in the Figure 6. Among the four videos, the first video (Vid As seen from Table 1, with the proper light intensity, the performance of vehicle detection
1) contains noisy video, which is mostly because of the dust on the camera. The second and third video (Vid 2 and count is most effective and at the night time, it is least effective. There are several DRR
and Vid 3) are taken from a newly installed camera at 3PM and 5PM in a single day. The intensity of light at 3PM factors associated to the inefficacy of stream at night time such as the black and white maintenance
is more than at 5PM, which is why they are included in or comparison. The fourth video (Vid 4) is taken in the (B/W) image stream, the light from the headlight of the vehicles, the switching from B/W to 33
night at 3PM. The four experimental videos provide a practical environment for comparison of how the quality of color and color to B/W during a sudden change in light intensity from vehicle’s headlight,
image affects in the performance of our system. The results of count in the four videos are shown in Table 2. etc. In overall, the overall accuracy (OA) among the four videos was 94%.