You are on page 1of 3

Abstract

Related works
For the recent years image processing and object detection are very common and applying
them to the smart parking place system can make the whole parking process less complex and
more efficient for both drivers and the manager.

In [Deni Kristin2020], to detect parked cars, this study uses the haar cascade classifier and
YOLOv3 model. Haar cascade classifier is an object detection method for computer vision, this
method was proposed by Paul Viola and Michael Jones in their paper in 2001. First the image is
converted into grayscale image and then several processes apply like Haar-like feature, Integral
image, Adaptive Booting and cascade classifier, then it will produce an object output that is
detected in the image.

Hình 0.1 [Deni Kristin2020] Haar cascade classifier

YOLO uses a CNN approach to detect objects in an image. A model is applied to the image,
then this network divides the image into several regions and predicts each boundary box for
each area, these bounding boxes then compared with each predicted probability.

Hình 0.2 [Deni Kristin2020] YOLOv3

In the result of the paper experiment, with the same condition to detect parker cars, the
average accuracy/scenario of the YOLOv3 has a better accuracy than the Haar cascade
classifier about 33%.
Authors in [yang-ding2019] have utilized YOLO (you only look once) v3, an end-to-end object
detection algorithm and apply it to detect free parking space and vehicles in parking lots. The
YOLOv3 uses the network of Darknet53 and contains 53 convolution layers. Darknet53 consists
of 5 residual blocks, each block consists of multiple residual units and each unit is constructed
by inputting residual operations with two DBL units. The DBL unit contains convolution, batch
normalization and leaky relu activation functions.

Hình 0.3 [yang-ding 2019] YOLOv3 basic model

Based on the structure of the YOLOv3, this paper adds a residual block to extract deep vehicle
parking space features and uses four different scale feature maps for object detection.

Hình 0.4 [yang-ding2019] Improved network structure

For the dataset used to train the network in this paper are picture of the parking lot of the
Pontifical Catholic University of Parana(PUCPR) in Brazil, the parking lot dataset PUCPR is
included in many different days and under different lighting conditions including sunny, rainy and
cloudy conditions. The results of the paper show that the change in the network structure
improves the accuracy of vehicle and parking space detection in the parking lot.

Hình 0.5 [yang-ding 2019] Comparision of the result of the two algorithms

However the lighting condition affects the output result pretty much. 

In [Raj Kirtibhai Patel 2021] to detect cars in the parking spot, the author chose Faster R-
CNN(two-stage) and YOLOv4(one stage). “faster_rcnn_R_50FPN_3x” implementation of Faster
R-CNN was used from Detectron2, whereas the YOLOv4 implementation came from Darknet.
To test and train 2 models this paper used CARPK dataset which contains drone-view images
of 90000 cars in total, and 2 models were trained until the accuracy on the test stopped
improving. In the result of car detection, the paper shows that the mean average precision(mAP)
of YOLOv4 performs much better than Faster R-CNN. The result of car tracking, the paper
shows that the YOLOv4 model has a 10% higher overall recall while the Faster R-CNN model
has 4% higher overall precision.

You might also like