IoT-Based Emergency Vehicle Detection Using YOLOv8

Researchers have been interested in the application of technology in emergency services. One such piece of technology that can help speed up emergency service response times is emergency vehicle detection systems. Systems for identifying on-road emergency vehicles make use of a variety of sensors, including cameras, radar, and GPS.

Baghel et al. [1] explore the effectiveness of the Ex-YOLO algorithm for detecting emergency vehicles compared to other real-time algorithms. The authors then propose an improved version of YOLO, called Ex-YOLO, which uses convolutional neural networks to improve accuracy. The authors compare Ex-YOLO with real-time algorithms like Faster R-CNN and SSD for emergency vehicle detection. The findings indicate that the Ex-YOLO algorithm performs better than other real-time algorithms in terms of accuracy and speed. The authors also discuss the limitations of their study, such as the limited size of their dataset. Overall, the paper presents a well-researched study on the effectiveness of the Ex-YOLO algorithm for emergency vehicle detection.

The research “Audio-Vision Emergency Vehicle Detection” proposes a new approach for emergency vehicle detection by integrating audio and visual information. The authors use a dataset of traffic videos with accompanying audio signals and contrast the performance of their proposed approach with traditional visual-only methods. The outcomes demonstrate that the proposed approach is better than traditional visual-only methods in terms of accuracy, recall, and precision. The paper thoroughly explains the proposed approach, including the audio feature extraction process, and utilizes deep neural networks for classification. The results suggest that such an approach can improve the accuracy and reliability of emergency vehicle detection systems, which can have significant implications for traffic safety and emergency response. However, it would be useful to evaluate the proposed method on a larger and more diverse dataset to further assess its effectiveness [2].

Managing traffic is a huge challenge on a global scale. traffic congestion problems are rising quickly, particularly in India, due to urbanization, population increase, and an increase in the number of automobiles. As a result, it is sometimes difficult for the ambulance to reach the hospital on time. Technological advancements have created numerous solutions to address serious problems and help save people.

The IoT surely has the potential to make the situation better. This study aims to examine several IoT movement control strategies and various methods for helping emergency vehicles arrive at nearby medical facilities in time. Finding the best possible methods to lessen traffic congestion is the key goal [3].

Most of the individuals who get hurt in auto accidents get help from other drivers or passengers. However, in case of a car accident that took place in a remote area or if the driver is the only one inside the vehicle and is unconscious, no one will be around to alert the appropriate authorities in time for medical treatment. A technique for identifying high-speed head-on and single-vehicle crashes, assessing the context, and sounding an alert is required in light of these problems. Chang, W. J [4] proposes the Internet of Vehicles (IoV) system named DeepCrash to address these issues. DeepCrash consists of an invehicle infotainment (IVI) telematics platform with a vehicle self-collision detection sensor and a front camera, a cloud-based deep learning server, and a cloud-based management platform. When a head-on or single-car collision is found, accident detection data is transferred to the cloud for self-collision vehicle accident recognition, and a corresponding emergency alert is sent out. According to the experimental findings, traffic collision detection accuracy may approach 96%, and the typical notification reaction time for emergencies is around 7s.

The desire for quicker and more precise detectors is increasing as autonomous vehicles and racing become increasingly popular. Even though even from great distances, our naked eyes can nearly quickly extract contextual information, picture quality and processing resource restrictions make recognizing tiny objects (that is, things that fill less than one square a constrained region of pixels in the input image), ajob that is difficult for machines and is an open topic for investigation The study [5] looks at ways to enhance the well-known YOLOv5 object detector to better identify up small things with a focus on autonomous racing. To do this, we look into how changing some of the model’s structural components (along with their links and various factors) might impact effectiveness and the time to get inferences. Consequently, we suggest a set of models of varying sizes that we call “YOLO-Z.” When recognizing smaller objects at 50% IOU for just 3ms longer inference time than the original YOLOv5 model, these models demonstrate an improvement of up to 6.9% in mAP. The objective is to continue investigating the possibility of modifying a well-known detector like YOLOv5 to handle certain jobs and offer perceptions on how particular alterations might affect small item identification in the future. Such findings may increase the amount of contextual data that autonomous vehicle systems have access to when they are applied to a larger context.

Current object detection networks use region proposal techniques to predict the locations of objects. Innovations like SPPnet and Fast R-CNN have reduced the running time of these detection networks, exposing region proposal computation as a bottleneck. The area Proposal Network (RPN) was created in the current study to cooperate with the detection network to exchange full-image convolutional features and provide almost-free area recommendations. An RPN, or fully convolutional network, predicts object bounds and objectness scores for each location simultaneously. By utilizing RPNs that have completed end-to-end training, Fast R-CNN conducts detection employing high-quality region suggestions. RPN and using just alternating optimization, a fast R-CNN may be trained to share convolutional features. This detection technique delivers cutting-edge item recognition with 300 proposals per image. The accuracy of PASCAL VOC 2007 (73.2% mAP) and 2012 (70.4% mAP) while operating ata frame rate of 5fps (including all phases) on a GPU [6].

The number of automobiles on the road will increase significantly every year. According to data from Malaysia’s road transport department (JPJ), as of December 31, 2019, there were around 31.2 million motor vehicle units registered in the country. As of the middle of 2017, Malaysia had a total of about 28.18 million motor vehicle units. Since traffic congestion may be identified by using the volume of cars as beneficial data, it is crucial to swiftly and effectively detect vehicles on the road. This is because it will help with traffic management. Muhammad et al. [7] use TensorFlow, a framework for machine learning, together with the object identification method YOLO, to construct deep learning for real-time car recognition. The suggested method in this study combines these two and other requirements using Python as the programming language to assess how well the YOLOv4 algorithm performs in comparison to the previous model in the vehicle identification system. To effectively count the number of cars passing in the video, this vehicle identification also employs the DeepSORT method. Yolov4, the top YOLO model from this research, achieved cutting-edge performance with 82.08% AP50 at a real-time frame rate of roughly 14 FPS on a GTX 1660ti. Yolov4 employed a custom dataset to achieve these results.

Detection of vehicles for aerial photos has become an important engineering approach and demonstrates the worth of academic study by using UAVs in intelligent transportation systems. This study provides a YOLO deep learning-based approach for identifying vehicles in aerial images. The method combines three open-source aerial picture datasets into one suitable for training the YOLO algorithm. Experiments show that the trained model matches the real-time requirements and works well on unknown aerial photographs, especially for small objects, rotating objects, and compact and dense objects [8].

One of the key facets of an economy is transportation. A large number of economic sectors struggle because of a loosely organized transportation network. This is an important challenge that emerging nations must deal with. There is no question that highways should be built to increase the throughput of the transportation network, but expanding alreadyexisting roads is also not practical in countries like Sri Lanka because of their shrinking geographical area and growing population. Hence, for us to resolve this problem, a more effective, technologically advanced approach must be embraced. The current pandemic turmoil has proven the need to prioritize ambulances when it is involved in a traffic congestion, in addition to the normal congestion scenarios. Another crucial aspect of the road system is pedestrians. Road accidents will be decreased through effective and safe crosswalks for pedestrians, which will also improve the current high traffic. The best option in this scenario is a sophisticated traffic monitoring system with integrated traffic light management. To detect cars and pedestrians and provide priority to emergency vehicles, this article suggests a method for intelligent, dynamic traffic monitoring and control. To achieve 91.3% detection precision, a new CNN is trained using the YOLOV3 architecture [9].

A computer vision-based method for real-time identification of several kinds of emergency vehicles in congested traffic is presented in this study. It enables the traffic controller to give emergency vehicles preferred path clearance, which could possibly save lives, safeguard property, and simplify the effort to prevent crimes. The proposed model is based on four classes: Firetrucks, Ambulances, Police Cars, and Normal Cars. The YOLO algorithm’s top layers were redesigned and retrained to get new learned weights, whereas the bottom levels remained locked. The model has demonstrated good outcomes and excellent metrics in recognizing and categorizing emergency vehicles and regular cars after retraining using the suggested modified YOLOv5. Using the mAP metric, police cars achieved 98%, 96% for fire trucks, 89% for ambulances, and 97% for normal cars [10].

This paper presents a lightweight version of the popular object detection algorithm YOLO, called YOLO-LITE, that is optimized for non-GPU computers. The authors propose several optimization methods to reduce the computational requirements of the original YOLO algorithm, such as substituting fully connected layers with convolutional layers and using a smaller input resolution. They also introduce a new approach for anchor box clustering to improve object detection accuracy. In an experimental comparison of YOLOLITE and other cutting-edge object identification algorithms, the article shows that YOLO-LITE performs as well as or better while operating in real-time on low-power devices. Overall, the research provides a practical and efficient method for real-time object identification on devices with limited resources [11].

This research suggests a new object detection technique named Tinier-YOLO, which is a compact version of the popular YOLO algorithm designed for real-time object detection in resource-constrained environments. The authors introduce several modifications to the YOLO architecture, such as reducing the convolutional filters and the dimensions of anchor boxes, to achieve a smaller model size and faster inference speed. They also propose a new data augmentation technique to increase the stability of the model. The article offers experimental findings from several datasets, demonstrating that Tinier-YOLO performs better than other leading object detection methods while running at real-time speeds on low-power devices. Overall, the paper offers an efficient and effective solution for real-time object detection in constrained environments [12].

This study suggests a novel object detection approach for landslide identification in satellite remote sensing pictures. The authors introduce a small attentional YOLO (You Only Look Once) model that improves the precision of item recognition in complicated backdrops by using an attention method. The model uses several convolutional layers and attention modules to reduce the computational requirements and achieve real-time performance. The suggested method surpasses existing cutting-edge object detection techniques in terms of detection accuracy and computing efficiency, as shown by experimental findings on a landslip detection dataset in the study. The proposed approach has the potential to assist in early warning and mitigation of landslide disasters, providing a valuable contribution to the field of disaster management [13].

The proposed approach helps improve the accuracy of traffic sign identification using YOLO v4 and artificial training data generated by numerous GANs. The authors trained their proposed model on an available traffic sign dataset and assessed it on both the same dataset and a real-world dataset. The testing findings demonstrated that their strategy outperformed a number of cutting-edge techniques in terms of accuracy. The authors also conducted ablation studies to investigate the effectiveness of using synthetic training data generated by different GANs. The paper concludes that their proposed approach using YOLO v4 and synthetic training data generated by various GANs can significantly improve the accuracy of traffic sign recognition, which is crucial for the safety of autonomous vehicles [14].

The paper recommends an improved version of the YOLO network for real-time vehicle detection in embedded systems. The suggested strategy alters the network’s topology to keep detection accuracy while lowering computational complexity. The modified network is then implemented on an embedded system using a heterogeneous computing platform, which includes an FPGA and a CPU. According to experimental findings, the suggested technique detects vehicles with high accuracy, using 30% less processing time than the original YOLO network.

Figure 1.

YOLO architecture

The suggested method surpasses other modern object identification techniques in the research when compared to both detection accuracy and processing speed. The outcomes show that the suggested approach is appropriate for embedded systems’ realtime vehicle detection applications [15].

The Experiment was conducted on a 2km stretch of a road that is known to have frequent emergency vehicle traffic. The Pi Camera was mounted at a height of 5 meters, facing the road. The microcontroller, Raspberry Pi was connected to a cellular network to transmit data to a cloud server.

Our system consists of a camera mounted on a pole, a microcontroller, and a cloud server. The camera captures images of the road, and the microcontroller processes the images and sends them to the cloud. This server runs the YOLOv8 object detection algorithm to detect emergency vehicles in the images.

The dataset is obtained by live capturing of Emergency Vehicles on the Road. While training our YOLOv8 model, the dataset was divided into train, valid, and test sets in a ratio of 70:15:15. Our model is trained to detect four classes of vehicles: ambulances, Police cars, Fire Engines, and and VIP Vehicles. Figure 3 shows the sample dataset images used for experimental purposes.

Figure 3.

Sample dataset

The dataset was preprocessed before it underwent training using the YOLOv8 algorithm. The following steps were included as a part of preprocessing a) Resizing images: The images were resized to 416 X 416 pixels so that all the images in the dataset have equal dimensions. This not only speeds up the training process but also improves our model’s accuracy. OpenCV was used to resize the images. b) Converting images to YOLO format: The YOLO algorithm demands a specific format for the dataset. It is crucial to transform our raw data into a format that the YOLO algorithm can use in order to create a YOLO model. The prepared dataset must include Image data and label data. c) Image data: In common, the Image data can be in any image format like JPEG or PNG. It is just like a matrix that represents the pixels of the input image. Here, our data has 416 X 416 pixels. d) Label data: The label data contains the annotations for each image, which typically includes bounding box coordinates and class labels. The following class labels were given: Ambulance, Fire Engine, Police, VIP. The label data will be stored in a text file having the exact name as the corresponding image file, and having a “.txt” extension. Each row in the label file represents one object in the image and has the following information: Object Class Index, Object Center Coordinates, Object width, and height. This dataset was prepared using Roboflow. Roboflow is a cloud-based software platform that provides tools for creating and managing custom object detection datasets, including datasets for YOLO models. Some of the features of Roboflow include Data Augmentation, Labelling tools, Dataset management, Export to YOLO format, and Integration with deep learning frameworks. The dataset was distributed into Train, Valid, and Test sets in a ratio of 70:15:15.

YOLO is an object detection model, and the most recent addition to the family of YOLO algorithms is YOLOv8. The YOLOv8 algorithm is used to train our model. YOLOv8 uses DarkNet-53 as its CNN, which is 53 layers deep. It gives more precise results and is more efficient than previous YOLO versions and other object detection algorithms. The proposed emergency vehicle detection system uses the YOLOV8 algorithm to detect emergency vehicles in the video stream. This system consists of two components: the training component and the detection component. The training part uses the YOLOv8 algorithm on a dataset of emergency Vehicles. The detection module detects the emergency vehicles in real-time using a trained YOLOv8 algorithm. The pre-trained YOLOv8 model was fine-tuned on our emergency vehicle dataset. The Adam optimizer was used to train the model.

Mathematical model: Mathematics behind YOLO It is necessary to understand the following steps: 1)

To understand the YOLO algorithm, one should know that the object class, along with its bounding box, is predicted.

The Prepared Dataset was trained using YOLOv8 algorithm using Darknet Framework. The Darknet Framework is an open-source framework. The model was trained for 200 epochs with a batch size of 64. The Rate of learning was changed to 0.001, and the Momentum was set to 0.9. The loss function used was YOLOv3 loss function. The trained algorithm was eventually evaluated on a set of validation images. The Evaluation Metrics used were Precision, F1 Score, and Recall.

The detection module was implemented using the YOLOv8 algorithm and the OpenCV library. This module takes a video stream as input and provides emergency vehicle detection in real time.

Figure 4.

Images captured on roads

Figure 5.

(a), (b), (c), (d) Sample dataset

The detected Vehicles are highlighted using bounding boxes and are hence labelled with their classes. Figure 6 shows a sample frame from the detected video stream. The detected emergency vehicles are highlighted with bounding boxes and labeled with their respective classes.

Figure 6.

Emergency vehicle identification using YOLOV8

The effectiveness of our system is measured by the mean average precision (mAP) metric. The efficiency of our system is compared against the YOLOv3 model and a baseline system that uses traditional computer vision techniques.

A confusion matrix is a performance evaluation tool that provides insights into how well a YOLO model is performing for object detection tasks.

A confusion matrix is a table that compares the YOLO model’s predicted labels to the test set’s true labels. Figure 7 shows the Confusion Matrix of the experimental setup. Here is what each category represents: True Positive (TP): The YOLO model correctly detected an object in the image. False Positive (FP): The YOLO model predicted an object in the image, but there was no object in the ground truth label. True Negative (TN): The YOLO model correctly predicted that there was no object in the image. False Negative (FN): The YOLO model failed to detect an object in the image that was present in the ground truth label. By analyzing the values in the confusion matrix, one can get insights about Precision and Recall, Accuracy, False Positive rate, and Error analysis.

Figure 7.

Confusion matrix

F1 Curve plots the F1 Score against a range of confidence thresholds for object detection in a YOLO model. The F1 score is a

This metric combines Recall and Precision and is used to assess how well the model is working. Figure 8 shows the F1 Confidence curve for vehicle identification, with a range of 0 to 1.

Figure 8.

F1 Confidence curve on the vehicle identification

The P curve represents a YOLO model’s object detection precision vs. recall plot. Precision and recall are two important indicators for assessing the effectiveness of object identification algorithms, and the P curve can provide several insights into a YOLO model’s performance. Figure 9 shows the Precision confidence curve for vehicle identification.

Figure 9.

Precision confidence curve for the vehicle identification

The PR (Precision-Recall) curve plots precision against recall for object detection in a YOLO model.

The recall-confidence curve plots recall versus confidence for object detection in a YOLO model. While confidence is the likelihood that an object will actually be found within a projected bounding box, recall is the proportion of the actual objects that the model effectively detects. Figure 10 shows the Precision-Recall curve for different emergency vehicle identification.

Figure 10.

Precision-Recall curve for different emergency vehicle identification

Figure 11.

Recall-Confidence curve on different vehicles

We used the SMTP server of IoT to deliver email notifications. Email clients use SMTP to send and receive email messages. In the context of IoT, an SMTP server can be used to send alerts, notifications, and other information via email.