With the development of new technology, the urban infrastructure, which is necessary to meet the social, economic, and physical needs of the population, is also gradually improving. Therefore, cities face significant urban planning challenges. With the increasing population and number of cars on the road parking space occupancy detection techniques are of great importance for eficient parking space management. Knowing the availability of parking spaces within a city in real-time and informing residents can be of great help in reducing the number of car park councils, reducing trafic congestion in cities, or improving the scalability of car parks and the time needed to find empty parking space. In many car parks, ground sensors are used to determine occupancy status. However, this requires the installation and maintenance of these sensors in every parking space. This can be expensive, especially in car parks with a large number of available spaces. With increasing computing power, new approaches to parking space occupancy detection using computer vision are emerging. However, these approaches are tailored to the site of application. In this work, we propose a comparison of two general approaches: classification and object detection. The reliability, precision, and scalability of these approaches are compared using a REST-API server application implemented in Python. In order to get as close as possible to general use, we use docker, which enables us easier develop and deploy this application independently of the hardware.
Parking in cities can sometimes be dificult or even impossible due to the increasing number of
cars. The number of cars on the road is increasing more and more with every year. Some parking
lots are full and others are almost empty. This inefective management and not informing the
public of parking lot occupancy information often leads to trafic problems such as congestion.
People often struggle to find empty space and spent a lot of time driving around. The solution
could be provided by smart parking lots, that could be able to inform in real-time about their
occupancy status. There are many parking lots, with occupancy detection using sensors [
For parking lot occupancy detection are existing several methods. One group can be described as detection with the use of some sensors and a second group that is using machine learning and computer vision.
Nowadays the most used approach to detecting the occupancy of parking spaces is by using sensors. This approach relies on the use of diferent types of sensors, which are located above or directly in each parking space.
The occupancy detection approach using sensor-based parking spaces is very accurate and
reliable, but a major disadvantage of these solutions is the cost required to build such a solution,
as each parking space needs its own sensor to detect occupancy. Although there are solutions
where a single tracking device can be used for multiple parking spaces [
Advances in deep learning have revolutionized machine learning by delivering unmatched
performance and methods for automatically extracting properties from images without the need
for human intervention. This allows the computer to quickly process video input to classify it
or learn to find objects [
The use of computer vision methods in the field of intelligent car parks makes it possible
to create less costly solutions with the help of several cameras and computing equipment,
and at the same time, they allow you to view the current situation in the car park from these
cameras [
Segmentation and classification In segmentation and classification, the processed image is first divided into smaller images called patches based on the segmentation mask. The segmentation mask contains the coordinates of the individual parking spaces. An illustration of a segmentation mask and patches can be seen in the figure 1.
After segmentation, each patch is processed using the selected machine learning algorithm and classified into specific classes, e.g. occupied or free.
Object detection allows us to identify and locate objects in an image. With this approach whole image is processed by an object detection neural network. The output of this neural network is a list of bounding boxes with corresponding labels of objects that were found in the image. 1http://cnrpark.it 2https://www.pixevia.com/smart-parking-occupancy
The topic of smart parking spaces and occupancy detection is due to the growing number of vehicles in trafic is still a lively topic and therefore there is already more research dealing with diferent approaches to addressing this issue.
Since this method is widely used there are many studies that develop and investigate parking
occupancy detection solutions using sensors [
The architecture of these solutions consists of the use of several diferent devices. Peripheral devices consisting of an Arduino UNO microcontroller, which had two infrared proximity sensors and two ultrasonic sensors, were placed in the parking spaces themselves. The Arduino UNO has very low energy consumption, so it was possible to power them from a battery. These peripherals were tasked with collecting data from the sensors and sending them to central devices using Bluetooth Low Energy. In this solution, the peripheral device was able to monitor up to 8 parking spaces. The central devices were Raspberry Pi microcomputers, which stored sensor data in an SQLite3 database running on them. These devices then sent the collected data to the back-end using an Internet connection. This architecture can be seen in the following ifgure 3.
Peripheral data had to be collected periodically to reduce energy consumption. Therefore, it was necessary to implement communication between the peripheral devices and the central device at certain intervals. The central device then communicated with the server on the backend using the Rest API protocol.
This technique [
This system consists of four parts: data preprocessing, low-level feature extraction, multi-class SVM recognition, and MRF-based correction. The input image is first pre-processed and divided into small samples, each of which contains 3 parking spaces. The Gaussian model is then used to obtain the probability of the earth occurring in pixels in the samples. These probabilities are further used as input values for a trained multi-class SVM classifier, which analyzes and classifies samples into 8 parking condition classes. Finally, Markov’s Random Field (MRF) is used to resolve conflicts between two neighboring sites in order to improve recognition.
In this work[
The input video uses built-in hardware to connect directly to an existing street camera system. Detection includes high-performance deep learning algorithms that detect pedestrians and vehicles. Vehicles and parking spaces are delimited from the sequential image using the YOLO model, and their overlapping areas are then calculated, which are used to identify the occupancy status. Street lamps are automatically switched to low light or high light levels depending on the detection results. All this functionality is handled by the Jetson TX2 microcomputer.
Some proposed methods [
To resolve perspective errors, a YOLO detection method with object overlap identification has been proposed to improve occupancy detection results.
Computer vision is a powerful tool for detecting the occupancy of parking areas, but in order to create a suitable model with reliable detection, we must address the problems that occur in image processing, which can cause inaccuracies.
These problems include changing ambient light conditions: diferent intensities of sunlight during the day, street lighting, and weather: sunny, rainy, snow, which afect the brightness and colors of the image. Other problems may be caused by occlusions in the image. Here we can include, for example, trees, street lights, open parts of the vehicle, a poorly parked vehicle, or its shadow. The camera itself must also be taken into account, such as its location: height, angle, or optical properties, which strongly afect the image quality.
A big problem with many existing solutions is that they are tailored to the environment. In our work, we try to create a generally applicable tool for detecting the occupancy of parking spaces.
The goal is to create a REST-API server application that could be located independently outside a specific parking lot and at the same time would not be tied to only one parking lot. Parking lots that would like to use occupancy detection services then only send image information using HTTP requests to a specified server endpoint. This application then processes the video inputs and sends a response with the detection result. At the same time, it stores these detection results in a database.
The detection results can then be used to inform the public or smart cities through another endpoint or graphical user interface.
In the application, we want to implement resource usage logging for future comparison of two occupancy detection methods.
The application is implemented in containers in docker, which allows us to be independent
of hardware and easy development with deployment. The second advantage of dock and
containerization is the ability to scale resources and thus we can analyze how many resources
the application needs [
Docker image we have made consists of 3 separated containers: • Python Django Server - This container contains the main part of the server, all functional logic, and application occupancy detection logic • PostgreSQL - In this container is the database for the server for storing data about parking lots and occupancy. • Graphana - This container is intended for the implementation of a simple user interface for displaying data in a database.
Within the application, we use two basic approaches to detect occupancy, classification and object detection
For classification, we use a model with AlexNet architecture. This model showed excellent
results in the field of classification and computer vision. It was also used for classification on
the set CNRPark [
To detect objects in the image we use the Faster RCNN detector type with the ResNet50 FPN network model. It is a network built on the VGG-19 network, with the fact that it is enriched with residual connections between the internal blocks.
A pre-trained network is used, whose inputs and outputs are adapted to our needs. In this case, the number of inputs is the size of the input image and the number of detected objects is 2. The vehicles whose bounding boxes are found will be detected and compared with the positions of parking spaces to determine occupancy.
We have approached the problem of parking occupancy detection, which is a big problem in trafic management in big cities. Our solution, that we propose is a general parking lot occupancy detection approach using a decentralized rest-API server.
So far, we have implemented a working docker image with a functional CUDA-capable container for the Rest server. In the rest-API server, we have implemented all needed parking lot management models and occupancy detection using segmentation and classification. For classification, we have implemented endpoints for the training network model and for occupancy detection.
After implementing the second occupancy detection method using object detection we plan to compare those methods with experiments.
In this server application, we want to compare those two approaches using a set of experiments. We plan to do qualitative experiments such as accuracy of detection on the same and diferent data sets and quantitative experiments to measure time for setting up, time that the application needs to process one detection, and average application response time. Also, we want to analyze the resource usage of these approaches so we will compare the amount of GPU and RAM that is used by the detection process.
The work reported here was supported by the Scientific Grant Agency of Slovak Republic (VEGA) under grant No. VG 1/0759/19 and the Operational Programme Integrated Infrastructure for the project: Advancing University Capacity and Competence in Research, Development and Innovation (ACCORD), ITMS code 313021X329, co-funded by the European Regional Development Fund (ERDF) Grant Number: 313011W988.