3 Results and discussion
The YOLO object detection model was trained on 42 authentic traffic images of two
main junctions of the city Jabalpur. A total of five images were kept apart from the
training set for testing the results. Table
2
illustrates the results obtained on the test
images.
From the results in Table
2
, it can be concluded that YOLO gives us fair results
considering that we have a minimal dataset that consists only of 42 images. The mean
time taken for detecting the vehicles per image is 1.36 s. The time taken in object
detection mainly depends on the incorporated processing hardware. The above results
have been obtained on Intel(R) Xeon(R) CPU @ 2.00GHz processor. In Table
2
, the
results obtained for image no. 45 shows quite different results from the other four
images. It is an outlier. That image was not taken at a right angle, and it has many
vehicles overlapping each other. For correct predictions, it is necessary to take an
image that shows vehicles as discrete as possible, not overlapped by any other object,
as discussed in the system architecture section and shown in Fig.
3
(right portion
123
1214
H. Khan et al.
Fig. 8
Processing time for one image
Fig. 9
Actual object count
versus predicted object count
images). The time mentioned in Table
2
is calculated programmatically by taking
the difference between the time points just after giving the input image and the time
point just after getting output. However the time considered for illuminating the green
light was estimated to be 2 s. Figure
8
summarizes the reason behind taking 2 s before
turning the green lights ON. From the results in Table
2
, the meantime for vehicle
detection and counting is 1.36 s, which can go up to 1.5 s. In some cases, about 0.4 s
has taken in image preprocessing before the vehicle detection process. In the end, the
green light time calculation can take a maximum time of 0.1 s. The whole process sums
up in 1 s. Figure
9
represents the actual object count versus the predicted count. From
the figure it is evident that, the dark blue line curve (predicted) follows the orange
cure (actual) except for image number 45, reason behind such behavior has already
been described earlier. In Fig.
10
, the graph summarizes the accuracy of images based
on the object predicted in the image. As it is visible, image 31 has highest accuracy
while image 45 has the lowest accuracy. Thus the average estimated accuracy of the
adopted method was 81.1%.
The processing time of five images has been depicted in Fig.
11
. Here minimum
processing time is 1.25 s, whereas the maximum processing time is 1.76 s. The mean
processing time is approximately 1.3 s. From the above investigations it can be inferred
that the proposed intelligent traffic management system based on the single image
processing is self adaptive, highly accurate, fast and has the potential to be implemented
in the traffic clearance at the junctions.
123
Machine learning driven intelligent and self adaptive…
1215
Do'stlaringiz bilan baham: |