Detection of vacant parking space in adverse weather Based on Improved YOLO Network Model

doi:10.21203/rs.3.rs-4175314/v1

Download PDF

Research Article

Detection of vacant parking space in adverse weather Based on Improved YOLO Network Model

https://doi.org/10.21203/rs.3.rs-4175314/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The increasing urban population in Delhi has exacerbated the prevailing parking challenges, and traffic congestion. In response, governmental efforts to implement "Smart Parking" initiatives have been initiated; however, existing solutions encounter constraints. This paper presents a pioneering Optimal Parking Allocation Model (OPSAM) leveraging the YOLO (You Only Look Once) detection algorithm to revolutionize parking management in Delhi.

OPSAM signifies a paradigmatic shift from driver-side parking search to system-side allocation, thereby optimizing parking space allocation. Benefiting from YOLO's rapid detection speed and suitability for real-time object detection, this study employs the YOLO v3 network for vehicle and the occupancy status of parking space detection in parking lots. Experimental findings demonstrate the method's efficacy in enhancing vehicle and parking space detection accuracy while mitigating neglected detection rates.

Deep neural networks

object detection

parking space detection

YOLOv3

computational efficiency

outdoor parking scenarios

India's dynamic economic growth has precipitated a substantial expansion in the automotive sector, marked by a significant increase in vehicle registrations from 2000 to 2010, with continued upward momentum in subsequent years. According to the Economic Survey report of 2023, New Delhi, India, boasts an astounding 12.25 million motor vehicles. However, this vehicular proliferation has engendered a critical challenge – the scarcity of parking infrastructure. Alarmingly, New Delhi faces a deficit of 12.15 millions parking spaces when compared against the ratio of available parking spots to vehicle ownership. Consequently, the glaring gap between parking supply and the escalating demand exacerbates the urban parking problem. Many techniques like Agent based, sensor based, VMS, etc. are used in Delhi but all have limitations. Recently, sensor technology is used in Palika Bazar Parking, New Delhi. Traditional methods of parking space detection typically rely on different type of sensors, technologies such as ultrasonic, infrared ray and geomagnetic and infrared ray [1–4]. However, deploying and maintaining sensors in every single parking space, particularly in large parking facilities, can be prohibitively expensive. While these methods offer high accuracy, their cost implications render them less feasible for widespread implementation. To improve parking space detection, this paper proposes which is based on vision-based technology- YOLO v3, an object detection based model using a security camera assistant system. The system detects parking spaces occupancy status and vehicles with security cameras to reduce the searching time required to find vacant parking spaces. This model aims to enhance vacancy detection in real-time without compromising speed, thereby offering a viable alternative to traditional sensor-based approaches.

The paper is structured as: section 2 provides a comprehensive literature review highlighting existing approaches and their limitations in addressing parking space detection challenges. Section 3 presents the proposed model in detail, explaining the architecture and underlying principles of the YOLO v3-based object detection approach tailored for parking space detection in adverse weather conditions. Section 4 outlines the data sources utilized in the experimentation phase and presents the experimental results obtained through the application of proposed method. This section offers insights into the performance and efficacy of the model under various adverse weather scenarios. Finally, Section 5 concludes the study by summarizing the research findings and discussing avenues for future research endeavors aimed at further enhancing parking space detection systems in adverse weather conditions.

Object detection serves as a fundamental task in computer vision, involving the identification of specific features within images [5]. The author employs a multi size symmetric search window to extract symmetric regions, validated by an Ada-boost classifier [6]. Teutsch et.al [7] utilize sliding windows to identify candidate vehicles from aerial cameras, subsequently verified using an Ada-Boost classifier based on motion features.

Deep learning-based target detection methods have gained prominence in recent years. Simonyan et.al [8] proposes very deep convolutional neural network (CNN) architecture, known as VGG, for object classification. The Region Proposal-Convolutional Neural Network (RCNN) [9] integrates region proposal and convolutional networks, representing a significant milestone in applying deep learning to conventional target detection assignments. The author introduces Faster RCNN, which improves upon RCNN by integrating region proposal, classification, and regression into a single network, thereby enhancing detection speed [10]. YOLO (You Only Look Once) [11] innovatively combines target localization and recognition, achieving notable improvements in detection speed. YOLOv3 [12] extends this approach by utilizing multi-scale features, achieving significant performance gains with a mean Average Precision (mAP) of 57.9% on VOC datasets within 51 milliseconds. YOLOv3 excels in balancing accuracy and detection speed, making it a preferred choice for object detection tasks.

In the domain of parking lot management, detecting vacant parking spaces presents a significant challenge. Ahrnbom et al. [13] extract color and gradient size features in the LUV color space, employing a SVM (Support Vector Machine) based classifier to classify parking spaces into occupied or vacant states. The another author utilize deep CNNs to detect the occupancy status of parking spaces based on Local Binary Pattern (LBP) features [14]. The author, Thomas et al. [15] constructs a binary classifier CNN to discern parking space occupancy. Peng [16] proposes three novel features for occupancy status determination, including color characteristics of vehicle, corner features, and local gray-scale variation trained using a deep neural network. Systems such as those by Amato et al. [14] periodically capture images of parking lots, employing a manually constructed mask to identify parking spaces. However, manual creation of masks for each parking space across diverse parking lots poses a practical limitation.

Methodology

Our proposed method involves adapting the YOLOv3 object detection model to handle all weather conditions effectively.

1. Yolo v3 network structure

The YOLOv3 architecture is rooted in the Darknet53 network structure, comprising 53 convolutional layers. Inspired by the ResNet neural network [12], Darknet53 is structured into five residual blocks, each comprising multiple residual units. A residual unit is composed of residual operations fed into two Double (DBL) units. Notably, the DBL unit encompasses convolutional, batch normalization, and leaky ReLU activation functions [9]. Introducing residual units enables deeper network architectures, mitigating the risk of vanishing gradients.

2. Improved network structure

This paper extends the YOLOv3 architecture by incorporating residual blocks to enhance the extraction of intricate car parking space features. Additionally, it utilizes four distinct scale feature maps to facilitate vehicle detection, enabling deeper networks to capture finer details. In the YOLOv3 network, six Double (DBL) units and convolution of 1x1 are employed in the target detection output-layer. To mitigate the risk of vanishing gradients resulting from the reuse of augmented features, this study modifies the configuration of the DBL units. Specifically, six DBL units are transformed into two DBL units and two ResNet units. The improved YOLO v3 network structure is depicted in Fig. 1.

The experimental dataset employed in this research originates from a parking lot dataset, the PKLot dataset, the COCO dataset, PASCAL VOC dataset. The COCO and PASCAL VOC datasets are widely recognized datasets for object detection tasks, containing 70 and 30 object categories, respectively. For this study, only the category relating to “cars” is utilized. The PKLot dataset [17] specifically focuses on parking space classification and provides an extensive image dataset captured at the parking lot of the New Delhi Municipal Corporation (NDMCPL) in New Delhi. Each image in the PKLot dataset is accompanied by an annotation file, facilitating precise object localization and classification. This dataset incorporates diverse scenarios, including sunny days, rainy days, and cloudy days, introducing variations in outdoor empty parking space visibility due to varying illumination levels. To validate the experiment's robustness, vacant parking spaces under all weather conditions are utilized as both the training sets and test sets.

On rainy and cloudy days, the dataset comprises 1,041 occupied parking spaces and 2,553 vacant parking spaces. Sample images depicting diverse scenarios are illustrated in Figs. 2 and 3. For sunny day scenario within the dataset, there are 16,430 occupied parking spaces and 14,272 vacant parking spaces. Introducing obstruction to the dataset results in 6,986 occupied parking spaces and 15,076 vacant parking spaces. This dataset selection allows for a comprehensive evaluation of the proposed model's performance across diverse weather conditions and obstruction scenarios.

The data-set comprises a total of 9,475 images, with detailed characteristics summarized in Table 1.

Table 1

Overview of the characteristics of the NDMCPL subsets
Parking	Time	Weather	No. of days	No. of images
Parking	Time	Weather	No. of days	No. of images	Occupied Percentage	Vacant Percentage
NDMCPL (300 parking spaces)	Day time	Sunny	58	3256	79.42%	20.58%
		Rainy	27	750	68.1%	31.9%
		Overcast	39	1880	45.63%	54.37%
	Night time	Sunny	45	2360	56.35%	43.65%
		Rainy	19	350	43.19%	56.81%
		Overcast	27	879	35.79%	64.21%

Data Set Processing

The processing of the PKLot dataset necessitates conversion into a format compatible with the PASCAL VOC dataset structure. Initially, the coordinate information contained within the XML files is extracted and reformatted to align with the specifications of the VOC dataset. Despite YOLOv3's commendable performance in target detection tasks using daily detection datasets, further refinement is essential to adapt it for parking lot detection tasks. The improved YOLOv3 algorithm model integrates a feature pyramid network structure, merging and linking feature maps from different levels to produce four sets of predictive feature maps. Following this, positional and class predictions are performed on these four sets of predictive feature maps.

Experimental Results

The experimental hardware setup for this study includes an Intel(R) Gold 5218R CPU @ 2.2 GHz, equipped with an 8-core CPU, 256GB memory, and a dedicated RTX 3090 graphics card. The operating system utilized is Windows 8. The algorithm is developed using the PyTorch framework, with a specific focus on PyTorch version 1.7.0 and CUDA version 11.0.

In the conducted experiment, the specified parameters were set as follows: a batch size − 16, an initial learning rate of 0.001, and adjustment to 0.0001 after 50 epochs. The total training duration spanned 100 epochs. The training concluded automatically when the validation loss failed to decrease consistently, signifying convergence of the model. Data allocation for the experiment involved a division into 70% for the training set, 10% for the validation set, and 20% for the test set. Notably, the training-validation set constituted 80% of the overall dataset.

Analysis of Daytime Detection Algorithm Experimental Results

In order to assess the efficacy of the algorithm designed for all-day detection of outdoor parking spaces by YOLO v3, recognizing the distinct characteristics of nighttime scenes, and the slightly elevated image darkness.

To enhance the detection accuracy specifically for nighttime outdoor parking spaces, a test set comprising 4596 nighttime outdoor parking lot images is generated using OpenCV based on the parking lot dataset. This experimental validation demonstrates the effectiveness of knowledge refinement through YOLOv3 for the purpose of enhancing detection accuracy.

The investigation into outdoor parking lot detection algorithms represents a pivotal area of research, utilizing deep learning methodologies to address critical aspects of parking space detection. This research not only holds significant implications for unmanned driving applications but also intersects with vehicle volume recognition and automated parking systems. The focus of this paper has been the examination of outdoor parking spaces, with the goal of determining their occupancy status through intelligent image analysis. The proposed methodology can be seamlessly integrated into real-world scenarios, either through hardware systems or dedicated applications (APPs). In practical parking scenarios, drivers can leverage this intelligent system or APP to accurately ascertain available parking space information, eliminating the need for blind searches and thereby alleviating unnecessary traffic congestion in parking lots.

The experiment in this study employs a diverse parking lot dataset encompassing various weather conditions. By leveraging deep learning techniques, image targets are effectively detected, showcasing notable improvements in accuracy compared to the original, unimproved methods. The key contributions of this research can be summarized as follows:

Outdoor Parking Space Detection Methods: In response to the requirements of outdoor parking space detection, this research delves into the application of deep learning in outdoor parking lot scenarios. Through experimental comparisons and analyses, two distinct detection methods are proposed, addressing parking space detection during both day and night. The structural and algorithmic aspects of these methods are elucidated.
Small Target Handling: Recognizing the prevalence of small target parking spaces in outdoor parking lots, the research introduces a data enhancement technique to oversample image data. This approach generates more small targets for the model to learn, enriching the model's feature extraction capabilities. Additionally, the VGG16 backbone network in the SSD network is replaced with a residual network, significantly enhancing the accuracy of the improved SSD detection network compared to the original SSD and other detection networks.
All-Day Outdoor Parking Space Detection Model: Considering the inclusion of nighttime scenes in outdoor scenarios, a novel all-day outdoor parking space detection model is proposed. This model obviates the need for additional nighttime training datasets and achieves enhanced nighttime detection accuracy through model fusion. The fusion of the YOLO model trained with nighttime data and the SSD model via the Encoder-Decoder structure of SID enables comprehensive all-day parking space image detection. The experimental results affirm that the fusion model not only improves nighttime detection accuracy but also reduces computational effort, achieving model light weighting without sacrificing accuracy.

In conclusion, this research significantly contributes to the advancement of outdoor parking space detection algorithms, providing robust solutions for real-world applications and enhancing the efficiency and accuracy of parking space detection systems.

Author Contribution

Abdul Ahad conceived the research idea and formulated the research objectives. Dr. Farhan Kidwai conducted the literature review and contributed to the theoretical framework development. Abdul Ahad designed the methodology, collected and analyzed the data, and implemented the proposed algorithm. Dr. Farhan Kidwai assisted in data analysis and interpretation. Abdul Ahad drafted the manuscript and prepared figures and tables. Dr. Farhan Kidwai critically reviewed and provided feedback on the manuscript. All authors read and approved the final version of the manuscript.

Shao, Y., Chen, P., Cao, T.: A Grid Projection Method Based on Ultrasonic Sensor for Parking Space Detection. In: IGARSS 2018–2018 IEEE International Geoscience and Remote Sensing Symposium. Valencia. pp. 3378–3381. (2018)
Zhou, F., Li, Q.: Parking Guidance System Based on ZigBee and Geomagnetic Sensor Technology. In: 2014 13th International Symposium on Distributed Computing and Applications to Business, Engineering and Science. Xian Ning. pp. 268–271. (2014)
Chen, H., Huang, C., Lu, K.: Design of a non-processor OBU device for parking system based on infrared communication. In: 2017 IEEE International Conference on Consumer Electronics - Taiwan (ICCE-TW). Taipei. pp. 297–298. (2017)
Yuan, C., Qian, L.: Design of intelligent parking lot system based on wireless network. In: 2017 29th Chinese Control And Decision Conference (CCDC). Chongqing. pp. 3596–3601. (2017)
Ding, X., Yang, R.: Vehicle and parking space detection based on improved YOLO network model. Journal of Physics: Conference Series, 1325, 1–7. (2019). https://doi.org/10.1088/1742-6596/1325/1/012084
Teoh, S., Bräunl, T.: Symmetry-based monocular vehicle detection system. Mach. Vis. Appl. 23, 831–842 (2012)
Teutsch, M., Krüger, W.: Robust and fast detection of moving vehicles in aerial videos using sliding windows. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Boston. pp. 26–34. (2015)
Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. (2014). https://arxiv.org/abs/1409.1556
Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Computer Vision and Pattern Recognition (CVPR). Columbus. pp. 580–587. (2014)
Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. IEEE Trans. Pattern Anal. Mach. Intell. 39, 1137–1149 (2017)
Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You Only Look Once: Unified, Real- Time Object Detection. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas. pp. 779–788. (2016)
Redmon, J., Farhadi, A.: YOLO V3: An Incremental Improvement. IEEE Conference on Computer Vision and Pattern Recognition, 2018:89–95. (2018)
Ahrnbom, M., Åström, K., Nilsson, M.: Fast Classification of Empty and Occupied Parking Spaces Using Integral Channel Features. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Las Vegas. pp. 1609–1615. (2016)
Amato, G., Carrara, F., Falchi, F., Gennaro, C., Vairo, C.: Car parking occupancy detection using smart camera networks and Deep Learning. In: 2016 IEEE Symposium on Computers and Communication (ISCC). Messina. pp. 1212–1217. (2016)
Thomas, T., Bhatt, T.: Smart Car Parking System Using Convolutional Neural Network. In: 2018 International Conference on Inventive Research in Computing Applications (ICIRCA). Coimbatore. pp. 172–174. (2018)
Peng, C., Hsieh, J., Leu, S., Chuang, C.: Drone-Based Vacant Parking Space Detection. In: 2018 32nd International Conference on Advanced Information Networking and Applications Workshops (WAINA). Krakow. pp. 618–622. (2018)
de Almeida, P.R., Oliveira, L.S., Britto, A.S., Silva, E.J., Koerich, A.L.: PKLot - a robust dataset for parking lot classification. Expert Syst. Appl. 42, 4937–4949 (2015)

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Detection of vacant parking space in adverse weather Based on Improved YOLO Network Model

Status:

Version 1

Abstract

Figures

Introduction

Related Work

Methodology

1. Yolo v3 network structure

2. Improved network structure

Experimental Dataset and Conditions

Data Set Processing

Experimental Results

Analysis of Daytime Detection Algorithm Experimental Results

Conclusion

Declarations

Author Contribution

References

Additional Declarations

Status:

Version 1