Enhancement of convolutional neural network for urban environment parking space classification

Abstract

BACKGROUND AND OBJECTIVES: The increase in the number of vehicles has several negative impacts, including traffic congestion, air pollution, noise levels, and the availability of parking spaces. Drivers looking for parking spaces can cause traffic jams and air pollution. The solution offered at this time is the development of a smart parking system to overcome these problems. The smart parking system offers a parking availability information feature in a parking area to break up congestion in the parking space. Deep learning is a successful method to solve parking space classification problems. It is known that this method requires a large computational process. Th aims of this study are to modified the architecture of Convolutional Neural Networks, part of deep learning to classify parking spaces. Modification of the Convolutional Neural Networks architecture is assumed to increase the work efficiency of the smart parking system in processing parking availability information.METHODS: Research is focusing on developing parking space classification techniques using camera sensors due to the rapid advancement of technology and algorithms in computer vision. The input image has 3x3 dimensions. The first convolution layer accepts the input image and converts it into 56x56 dimensions. The second convolution layer is composed in the same way as the first layer with dimensions of 25x25. The third convolution layer employs a 3 x 3 filter matrix with padding of up to 15 and converts it into 10x10 dimensions. The fourth layer is composed in the same way as the third layer, but with the addition of maximum pooling. The software used in the test is Python with a Python framework.FINDINGS: The proposed architecture is the Efficient Parking Network or EfficientParkingNet. It can be shown that this architecture is more efficient in classifying parking spaces compared to some other architectures, such as the mini–Alex Network (mAlexnet) and the Grassmannian Deep Stacking Network with Illumination Correction (GDSN-IC). EfficientParkingNet has not been able to pass the accuracy of Yolo Mobile Network (Yolo+MobileNet). Furthermore, Yolo+MobileNet has so many parameters that it cannot be used on low computing devices. Selection of EfficientParkingNet as a lightweight architecture tailored to the needs of use. EfficientParkingNet's lightweight computing architecture can increase the speed of information on parking availability to users.CONCLUSION: EfficientParkingNet is more efficient in determining the availability of parking spaces compared to mAlexnet, but still cannot match Yolo+MobileNet. Based on the number of parameters, EfficientParkingNet uses half of the number of parameters of mAlexnet and is much smaller than Yolo+MobileNet. EfficientParkingNet has an accuracy rate of 98.44% for the National Research Council parking dataset and higher than other architectures. EfficientParkingNet is suitable for use in parking systems with low computing devices such as the Raspberry Pi because of the small number of parameters.

Keywords

• deep learning
• efficient parking network
• malexnet
• mobilenet, parking space