IOT Based Intrusion Detection

IOT Based Intrusion Detection — Topics for IoT Students

FireNet: A Specialized Lightweight Fire & Smoke

Detection Model for Real-Time IoT Applications Arpit Jadon, Student Member, IEEE Mohd. Omama Akshay Varshney Z.H. College of Engg. and Tech.

Z.H. College of Engg. and Tech. Z.H.

College of Engg. and Tech. Aligarh Muslim University Aligarh Muslim University Aligarh Muslim University Aligarh, India Aligarh, India Aligarh, India arpitjadon@zhcet.ac.in mohd.omama@gmail.com akshayvarshney.001@gmail.com

Mohammad Samar Ansari, Member, IEEE Rishabh Sharma

Software Research Institute Z.H. College of Engg. and Tech. arXiv:1905.11922v2 [cs.CV] 4 Sep 2019

Athlone Institute of Technology Aligarh Muslim University

Ireland Aligarh, India mansari@ait.ie rishabh.sharma079@gmail.com

Abstract—Fire disasters typically result in lot of loss to life detection systems need a sufficient level of fire initiation for a and property. It is therefore imperative that precise, fast, and clear detection, which leads to a long detection delay causing possibly portable solutions to detect fire be made readily available irreparable damages. An alternative, which could lead to the to the masses at reasonable prices.

There have been several research attempts to design effective and appropriately priced enhancement of robustness and reliability in the present fire fire detection systems with varying degrees of success. However, detection systems, is the visual fire detection approach. most of them demonstrate a trade-off between performance and As a courtesy of the advancement in various artificial intel- model size (which decides the model’s ability to be installed ligence fields, vision-based research fields like Image Process- on portable devices).

The work presented in this paper is an ing and Computer Vision have witnessed a fair share of fruitful attempt to deal with both the performance and model size issues in one design. Toward that end, a ‘designed-from-scratch’ benefits. Various deep learning (DL) [1] models have been neural network, named FireNet, is proposed which is worthy able to comfortably surpass the human level performance in on both the counts: (i) it has better performance than existing specific computer vision applications like image classification. counterparts, and (ii) it is lightweight enough to be deploy-able on The visual-based fire detection approach also has been able to embedded platforms like Raspberry Pi.

Performance evaluations take advantage of these technological improvements. Visual- on a standard dataset, as well as our own newly introduced custom-compiled fire dataset, are extremely encouraging. based fire detection systems have many advantages over the Index Terms—Convolutional Neural Networks, Embedded Sys- hardware-based alarm systems in terms of cost, accuracy, ro- tems, Fire Detection, Internet of Things (IoT), Neural Networks, bustness, and reliability.

Over time the handcrafted visual fire Smoke Detection. detection approaches, which offer lower performance in terms of accuracy and false triggering, have been replaced by deep I. I NTRODUCTION learning based approaches, which are better in performance in Due to the rapid increase in the number of fire accidents, terms of varying metrics. This better performance is attributed the fire alarm systems form an integral part of the necessary to the capability of the deep learning based approaches to accessories in any sort of construction.

Fire accidents are automatically extract features from the raw images. On the the most commonly occurring disasters nowadays. In order contrary, the handcrafted approaches require more careful to mitigate the number of fire accidents, a large number of handling as the features are to be extracted manually from methods have been proposed for early fire detection to reduce the input images.

Thus, by combining these more reliable the damage caused by such accidents. Apart from the problem visual based fire detection techniques with the conventional of early fire detection, present fire alarm systems also prove sensor based techniques, a more robust and reliable fire alarm to be inefficient in terms of the false triggering of the alarm system could be developed. Therefore, in this work, our aim systems.

Present fire detection methods are typically based is to introduce an advanced fire and smoke detection unit, on physical sensors like thermal detectors, smoke detectors, which is reliable, reduces the false triggering problem and and flame detectors. However, these sensor-based detection incorporates various state of the art technological concepts systems are not very reliable for fire detection. For instance, like Convolutional Neural Networks (CNN) and the Internet quite often there are cases of false triggering with smoke of Things (IoT) [2].

In order to get the best fire detection detectors, as it does not possess the capability to differentiate performance while maintaining a significantly good frame rate between fire and smoke. On the other hand, the other two on the Raspberry Pi 3B (1.2GHz Broadcom BCM2837 64bit CPU, 1 GB RAM computer [3]) during real-time fire detection,

we have developed our own neural network from scratch and This paper is organized as follows. Section II discusses the have trained it on a dataset compiled from multiple sources. past research work done in the area of deep learning and hand- The model is tested on a real world fire dataset, accumulated engineering based fire detection. Section III provides a detailed by us and also on a standard fire dataset provided by authors description of our dataset followed by the description of our in the work [4].

We have taken Raspberry Pi as the hardware proposed approach in section IV, which is in turn followed platform to deploy our model because it is the most cost- by the section V, which discusses the IoT implementation to effective platform for running computationally non-intensive develop a complete state of the art fire detection unit. Finally, deep learning algorithms and will serve well our purpose we discuss our results in section VI, followed by a discussion of developing a fire and smoke detection unit.

In addition, on the effectiveness of the proposed approach in Section VIII. the system is also capable of differentiating fire from smoke Section VIII contains concluding remarks. thereby reducing the false triggering problem by triggering distinct sound alarms for fire and smoke respectively. II. R ELATED W ORK However, smoke detection using vision-based techniques Detecting fire is an important issue, modern technology is in face many challenges.

Video processing techniques generally dire need of an appropriate detection system that can reduce work on the principle of reading pixel values of the color. the damage caused due to a large number of fire accidents Thus, in real time, it becomes difficult to distinguish between taking place everyday [7]. the greyish colored objects present in the image along with Initially, the researchers attempted to develop handcrafted smoke to be detected.

Moreover, smoke detection becomes techniques for fire detection by focusing on the motion and harder in a dark environment due to the camera’s capturing color properties of the flame detection. One such work done attributes [5]. Hence, for better overall performance, in our by Thou-Ho et al. [8] utilized both chromatic and dynamic work, we utilize a smoke sensor integrated with the system properties of fire and smoke for the true flame detection.

In for smoke detection, thus also removing the need for installing another work, Celik et al. [9] tried to distinguish fire from the separate fire and smoke detectors. It simultaneously minimizes smoke utilizing two different color spaces and in order to make the false triggering problem occurring with smoke detectors the classification more robust, they adopted concepts from used as fire detectors in conventional fire alarm systems. fuzzy logic to discriminate fire from the other fire like objects.

Smoke sensors are also economically affordable and can In contrast to [8], where they used RGB color space for flame detect smoke efficiently. To alert the end user about the fire detection, Turgay et al. [9] used the YCbCr color space and emergency, an IoT based remote data transmission system is made some modifications to overcome the drawback from the also employed to send MMS containing visual fire feedback previous technique by creating some more generic rules to and fire alert to the end user. detect the fire, but high false detection rate and restriction This kind of intelligent fire and smoke detection unit can be of detection only at a feeble distance were the associated used for a wide range of applications, such as giving an early drawbacks.

In addition to the color property, motion has also warning for an emergency, notifying fire brigades so that they been taken as the criterion to detect the fire in some works. can get to the site of fire as soon as possible, triggering the Rafiee et al. [10] used the static and dynamic properties of automatic fire suppression systems, etc. [6]. the fire and smoke. However, the false negative rate remains an issue here also due to the presence of other objects in the A.

Contributions background with similar color properties as the fire pixels. The main contributions of this work are as follows: Qiu et al. [11] proposed an auto-adaptive edge detection • We introduce a shallow neural network for fire detection, algorithm for flame detection. In another work, Rinsurongka- which unlike previous DL-based fire detection approaches wong et al. [12] used the dynamic properties of the fire for where bulky convolutional neural networks were used, flame detection, but this method also failed with images having can perform real-time fire detection at a frame rate that pseudo fire like objects in the background.

This drawback was surpasses the frame rates achieved until now. overcome by the authors of [13] in which they proposed two • We also introduce a new, small but very diverse, training optical flow estimators for differentiating fire from the non-fire fire dataset combining images from multiple sources. objects. Mobin et al. [14] introduced a fire detection system, Moreover, we also introduce another self-made dataset Safe from Fire (SFF) that uses multiple sensors to detect fire consisting of self-shot videos in a challenging environ- and smoke distinctly.

But the use of multiple sensors, caused ment. the system to be more expensive. • We also present a working implementation of a complete Although these hand-engineered approaches to fire detection fire detection unit that can suitably replace the conven- are not computationally expensive and can be deployed to tional physical sensor based fire detection alarm systems, economically feasible hardware like Raspberry Pi with a good simultaneously reducing the false and delayed triggering frame rate, there is the drawback of manual feature extrac- problems, along with providing a remote verification tion from the raw images.

This drawback makes the hand- functionality by providing real-time visual feedback in engineering task very tedious and inefficient, particularly when the form of an alert message using Internet of Things the number of images in the dataset is high. In contrast, the (IoT). DL-based approaches have the advantage of automatic feature

extraction, thus, making the process much more efficient and reliable than the conventional handcrafted image processing techniques. However, these deep learning approaches require a lot of heavy computational power, not only while training but also when the trained model is to be deployed to hardware for carrying out a specific task. In the case of fire detection, the capability of the algorithm to be deployed on computationally heavy hardware like a personal computer machine is futile because the unit needs to be comparable to a conventional fire detector, both in terms of physical size and cost.

Various deep learning approaches for fire detection have Fig. 1. Few images from our training dataset. been proposed. Zhang et al. [15] in their work on forest fire detection utilized fire patches detection with a fine-tuned pre- trained CNN, ‘AlexNet’ [16] while Sharma et al. [17] also training a network and expecting it to perform well in realistic proposed a CNN-based fire detection approach using VGG16 fire detection scenarios.

The reason that this dataset does not [18] and Resnet50 [19] as baseline architectures. But in both appear to be diverse enough is that it contains a large number these works, the large on-disk size and a number of parameters of similar images. Thus, we tried to create a diverse dataset by render these models unsuitable for on-field fire detection shooting fire and non-fire videos in a challenging environment, applications using low-cost low-performance hardware. and also by collecting fire and non-fire images with fire like Muhammad et al. fine-tuned different variants of CNNs objects in the background from the internet.

Our train dataset like AlexNet [20], SqueezeNet [21], GoogleNet [22], and consists of few fire and non-fire images sampled from the MobileNetV2 [23]. In [20]–[22], they used Foggia’s dataset Foggia’s and Sharma’s [17] dataset, and images taken from the [4] as the major portion of their train dataset, while in [23] internet (Google and Flickr). In order to maintain the diversity the train dataset was combined from [24] and [4].

Although in our train dataset, we also augmented the Sharma’s dataset Foggia’s dataset contains 14 fire and 17 non-fire videos with and randomly picked a few images from it. Thus, the final a large number of frames, the dataset contains a lot of similar train dataset consists of a total of 1,124 fire images and 1,301 images, which restricts the performance of the model trained non-fire images. Although the dataset may appear to be small, on this dataset to a very specific range of images.

Moreover, it is extremely diverse. in [23] the training dataset consists of 1844 fire images and For the test dataset, we tried to accumulate as many realistic 6188 non-fire images, which points towards an unbalanced images as possible because the fire detection unit has to dataset, and consequently the possibility of biased results. ultimately work in these situations only. However, since our More importantly, these networks have a large number of train dataset is already diverse enough, we used these realistic layers and have large on-disk size, which restricts their use images for testing purpose only. in embedded vision application.

The incorrect selection of Our complete test dataset consists of 46 fire videos (19,094 training dataset and bulky models with a large number of frames) with 16 non-fire videos (6,747 frames) and additional layers and parameters direct towards the need for a shallow 160 challenging non-fire images. To make our test dataset network, which has a good trade-off between fire detection diverse, out of all the frames extracted from this dataset, we accuracy and high frame rate, allowing the model to run randomly sampled few images from each video to form our efficiently on a low-cost embedded device.

Therefore, learning final test dataset to be used in this work. The model performed from our previous works, we tried to make our train dataset extremely well and the results are discussed in section VI. In more diverse and challenging by gathering images from Flickr Fig. 1 and Fig. 2, we have shown a few images from our and Google while combining these images with few images training and test dataset respectively. sampled from Foggia’s and Sharma et al. dataset [17].

To As a means to let the research community benefit from, tackle the problem of too big on-disk size and a large number and extend our efforts, we have open-sourced the dataset and of layers, we build our own neural network from scratch FireNet [25]. and name it as ‘FireNet’. Moreover, in order to help other researchers to build better fire detection models, we have open IV. PROPOSED APPROACH sourced our dataset and trained model.

More details about our The past deep learning based fire detection approaches like dataset and network are provided in the subsequent sections. [17], [20]–[23] only followed the process of fine-tuning differ- ent CNNs like VGG16, Resnet, GoogleNet, SqueezeNet, and III. DATASET DESCRIPTION MobileNetV2. One major drawback with just fine-tuning such We observed from the datasets used in the past approaches bulky CNNs is that the final on-disk size of the trained model [20]–[23] that currently there is a scarcity of a diverse fire and number of layers is too large, thus preventing these trained dataset.

One of the dataset provided by Foggia et al. [4] models to run smoothly at a sufficient frame rate on low- contains 31 fire and non-fire videos. Although Foggia’s dataset cost hardware like a Raspberry Pi for real-time fire detection. is vast, it is not diverse enough to be solely utilized for Also, it is quite obvious that the emphasis of the trained

Fig. 2. Few images from our test dataset.

model to run at a good frame rate on low-cost hardware like Raspberry Pi is very much valid, since the end goal of all these approaches is implementation for real-world applications, i.e., to be transformed into various fire detection units installed in the required environment like a shopping mall, a residence, hotel, etc. Thus, there is the need to use commercially available low-cost hardware, which is economically feasible unlike a high-cost extensive computational machine which is futile in real-world fire detection applications.

Therefore, we propose a light-weight neural network ar- chitecture called FireNet, that is suitable for mobile and embedded applications, which shows a favorable performance for real-time fire detection application. The network runs at a very satisfactory frame rate of more than 24 frames per second on less powerful, economically feasible single board computers like Raspberry Pi 3B, etc. The proposed neural network has three convolution layers and four dense layers (including an output ‘softmax’ layer).

Fig. 3. Architecture of the proposed neural network.

A. Architecture

B. Regularization The complete architecture diagram of the proposed network We have used dropout with convolution layers along with is shown in Fig. 3, from where it can be seen that FireNet the dense layers. The general reasoning trend is to use dropout contains a total of 14 layers (including pooling, dropout and with dense layers only.

However, we saw that the overall ‘Softmax’ output layer). There are three convolution layers, results of the neural network improved when dropout was used each of which are coupled with pooling and a dropout layer. with convolution layers. Hence, we opted for it.

We chose a Each of these layers has Rectified Linear Unit (Relu) as the standard dropout value of 0.5 for convolution layers. A dropout activation function except the last layer, which has Softmax as value of 0.2 was taken for the subsequent dense layer. This its activation function.

The total number of trainable parame- is because it has been shown that overfitting generally takes ters turns out to be 646,818 (size on disk ∼7.45 MB). place in initial layers of the neural network [26]. The first layer is a convolution layer, which takes a colored input image of size (64×64×3). This input size is selected V.

C OMPLETE F IRE D ETECTION U NIT AND I OT after empirical results that compared various sizes. The input I MPLEMENTATION size can be increased up to (128×128×3) without having a We deployed ‘FireNet’ to the Raspberry Pi 3B and inter- drastic effect on FPS. This layer has 16 filters with a kernel faced a smoke sensor and two distinct fire alarms to it, in order size of (3, 3). to detect fire and smoke distinctly and consequently triggering In each of the two subsequent convolution layers, we double the distinct fire alarm thus, eliminating the false triggering the input features keeping the kernel size constant.

This is problem associated with the smoke based conventional fire followed by a flatten layer and 2 dense layers of 256 and detectors or alarm systems. We also made the complete fire 128 neurons each. The final layer is a dense layer with two detection unit IoT capable, thus, allowing the development of neurons, acting as the output prediction layer. a completely autonomous unit with the potential of providing

TABLE I

T EST PERFORMANCE OF 'F IRE N ET'ON OUR REAL - WORLD TEST DATASET

Metrices Our dataset (%) Foggia’s dataset

(%) Accuracy 93.91 96.53 False Positives 1.95 1.23 False Negatives 4.13 2.25 Recall 94 97.46 Precision 97 95.54 F-measure 95 96.49

Fig. 4. Overview of the complete unit: (a) Camera (b) Raspberry Pi 3B

(c) Microcontroller (d) Cloud storage and SMS/MMS service (Amazon S3 and Twilio) (e) End-user device for receiving fire alert (visual and textual) (f) Smoke alarm (g) Fire alarm with a different sound than smoke alarm (h) Smoke sensor for sensing smoke and thus, aiding in fire-smoke differentiation.

visual fire feedback and alert message in the emergency situation. For implementing IoT based remote visual fire Fig. 5. Training and validation curves for model accuracy. feedback and alert, we utilized two cloud facilities, Twilio [27] and Amazon Web Service’s Simple Storage Service (AWS S3) [28].

Twilio is a messaging service that allows sending in Table I. We have used four different metrics (accuracy, SMS/MMS while AWS S3 is a file storage service. We utilize precision, recall, and F-measure) in order to present a complete AWS S3 to upload fire images or clips recorded at the time and reliable analysis.

As can be seen from Table I, the output of fire emergency to the cloud while using Twilio to send values obtained for each metric is pretty much up to the mark an MMS containing fire image or clip along with the fire on both datasets. Moreover, the performance of the model is alert message. Fig. 4 shows an overview of the complete IoT better on the Foggia’s dataset as the image diversity is limited enabled fire detection unit capable of differentiating between in this dataset.

In order to further ensure the reliability of the fire and smoke. The microcontroller has been used as an trained model, we have shown the training process using the analog to digital converter (ADC) for communicating the training curves. Fig. 5 represents the training and validation smoke sensor’s analog data to Raspberry Pi, as well as to loss curves obtained during the training process while Fig. 6 trigger the sound alarms.

In a future work, the microcontroller shows the training and validation accuracy curves obtained is expected to be done away with, by the inclusion of a special during the same process. As can be seen from Fig. 5 and purpose ADC for Raspberry Pi (e.g. MCP3008) in the design.

Fig. 6, the trained model is able to generalize significantly on VI. R ESULTS the validation set. Also, note that we have used a 70% and 30% split between the training and validation set.

Moreover, there is The aim of this work is to present a method that can be no overlap between the two respective sets. Another significant smoothly deployed to an embedded device in order to finally advantage of this network is that it can run on Raspberry Pi build a complete fire detection unit. Therefore, it becomes 3B at a frame rate of 24 frames per second. inevitable to use a test dataset that includes images that are often encountered in real-world fire emergencies with an image VII.

D ISCUSSION ON SIMILAR WORKS quality that is commonly obtained with a camera attached This section aims to draw a comparison between the pro- to low-cost hardware like Raspberry Pi 3B. Currently, there posed FireNet and other available state-of-art approaches for is an absence of such a fire dataset. Although the dataset fire detection.

The biggest advantage of FireNet is its small provided in work [4] contains such images but the dataset lacks size on disk (∼7.45 MB) which is attributed to the shallow diversity. However, for the sake of comparison we have shown network, and consequently the small number of trainable the results of our model on this dataset as well. Therefore, to parameters (646,818).

It needs to be acknowledged that there reflect the performance of our trained model in most realistic are better performing fire detection solutions available in situations, we compiled the test dataset from our own videos the literature. However, the definition of ‘better’ is largely shot in challenging actual world fire and non-fire environment. dependent upon the resources used and the dataset trained The performance obtained on both these datasets is shown on.

For instance, the works in [17], [20]–[23] all employed

[5] T.-H. Chen, Y.-H. Yin, S.-F.

Huang, and Y.-T. Ye, “The smoke detection for early fire-alarming system base on video processing,” in 2006 Inter- national Conference on Intelligent Information Hiding and Multimedia, pp. 427–430, IEEE, 2006. [6] “Applications for Fire Alarms and Fire Safety.” http://www.vent.co.uk/ fire-alarms/fire-alarm-applications.php. Accessed: 2019-01-05. [7] “Fire Accidents.” https://urlzs.com/7mTM.

Accessed: 2019-01-05. [8] T.-H. Chen, P.-H. Wu, and Y.-C.

Chiou, “An early fire-detection method based on image processing,” in 2004 International Conference on Image Processing, 2004. ICIP’04., vol. 3, pp. 1707–1710, IEEE, 2004. [9] T. Çelik, H.

Özkaramanlı, and H. Demirel, “Fire and smoke detection without sensors: Image processing based approach,” in 2007 15th European Signal Processing Conference, pp. 1794–1798, IEEE, 2007. [10] A. Rafiee, R.

Dianat, M. Jamshidi, R. Tavakoli, and S.

Abbaspour, “Fire and smoke detection using wavelet analysis and disorder characteristics,” in 2011 3rd International Conference on Computer Research and Development, vol. 3, pp. 262–265, IEEE, 2011. [11] T. Qiu, Y. Yan, and G.

Lu, “An autoadaptive edge-detection algorithm for Fig. 6. Training and validation curves for model loss. flame and fire image processing,” IEEE Transactions on instrumentation and measurement, vol. 61, no. 5, pp. 1486–1493, 2012. [12] S. Rinsurongkawong, M.

Ekpanyapong, and M. N. Dailey, “Fire detec- massive CNN based deep models which translate to large tion for early fire alarm based on optical flow video processing,” in 2012 9th International Conference on Electrical Engineering/Electronics, on-disk sizes, and typically reported detection capabilities of Computer, Telecommunications and Information Technology, pp. 1–4, around 4-5 frames per second while running on low-cost IEEE, 2012. embedded hardware.

FireNet on the other hand derives its [13] M. Mueller, P. Karasev, I.

Kolesov, and A. Tannenbaum, “Optical flow estimation for flame detection in videos,” IEEE Transactions on image impressive fire detection capabilities from being trained on processing, vol. 22, no. 7, pp. 2786–2797, 2013. a much more diverse dataset, as well as its specialized design [14] I. Mobin, M.

Abid-Ar-Rafi, M. N. Islam, and M.

R. Hasan, “An from scratch for use in fire detection. Using this effective intelligent fire detection and mitigation system safe from fire (sff),” Int.

J. Comput. Appl, vol. 133, no. 6, pp. 1–7, 2016. combination, FireNet is successfully able to provide real-time [15] Q.

Zhang, J. Xu, L. Xu, and H.

Guo, “Deep convolutional neural fire detection feature for upto 24 fps which is almost as good networks for forest fire detection,” in 2016 International Forum on Man- as human visual cognition. agement, Education and Information Technology Application, Atlantis Press, 2016. VIII. C ONCLUSION [16] A.

Krizhevsky, I. Sutskever, and G. E.

Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in neural infor- In this work, we present a very lightweight neural network mation processing systems, pp. 1097–1105, 2012. (‘FireNet’) build from scratch and trained on a very diverse [17] J. Sharma, O.-C.

Granmo, M. Goodwin, and J. T.

Fidje, “Deep convo- dataset. The ultimate aim of the complete work is to develop lutional neural networks for fire detection in images,” in International Conference on Engineering Applications of Neural Networks, pp. 183– an internet of things (IoT) capable fire detection unit that 193, Springer, 2017. can effectively replace the current physical sensor based fire [18] K. Simonyan and A.

Zisserman, “Very deep convolutional networks for detectors and also can reduce the associated problems of false large-scale image recognition,” 2014. [19] K. He, X. Zhang, S.

Ren, and J. Sun, “Deep residual learning for image and delayed triggering with such fire detectors. The introduced recognition,” in Proceedings of the IEEE conference on computer vision neural network can smoothly run on a low-cost embedded and pattern recognition, pp. 770–778, 2016. device like Raspberry Pi 3B at a frame rate of 24 frames per [20] K.

Muhammad, J. Ahmad, and S. W.

Baik, “Early fire detection using convolutional neural networks during surveillance for effective disaster second. The performance obtained by the model on a standard management,” Neurocomputing, vol. 288, pp. 30–42, 2018. fire dataset and a self-made test dataset (consisting challenging [21] K. Muhammad, J.

Ahmad, Z. Lv, P. Bellavista, P.

Yang, and S. W. real-world fire and non-fire images with image quality that Baik, “Efficient deep cnn-based fire detection and localization in video surveillance applications,” IEEE Transactions on Systems, Man, and is similar to the images captured by the camera attached to Cybernetics: Systems, no. 99, pp. 1–16, 2018. Raspberry Pi) in terms of accuracy, precision, recall, and [22] K.

Muhammad, J. Ahmad, I. Mehmood, S.

Rho, and S. W. Baik, F-measure is encouraging.

Moreover, the IoT functionality “Convolutional neural networks based fire detection in surveillance videos,” IEEE Access, vol. 6, pp. 18174–18183, 2018. allows the detection unit to provide real-time visual feedback [23] K. Muhammad, S. Khan, M.

Elhoseny, S. H. Ahmed, and S.

W. Baik, and fire alert in case of fire emergencies to the user. In our “Efficient fire detection for uncertain surveillance environment,” IEEE future work, we plan to improve the performance of the model Transactions on Industrial Informatics, 2019. [24] D.

Y. Chino, L. P.

Avalhais, J. F. Rodrigues, and A.

J. Traina, “Bowfire: on even a more diverse dataset. detection of fire in still images by integrating pixel color and texture R EFERENCES analysis,” in 2015 28th SIBGRAPI Conference on Graphics, Patterns and Images, pp. 95–102, IEEE, 2015. [1] Y. LeCun, Y.

Bengio, and G. Hinton, “Deep learning,” nature, vol. 521, [25] A. Jadon, M.

Omama, A. Varshney, M. S.

Ansari, and R. Sharma, no. 7553, p. 436, 2015. “Firenet-lightweight-network-for-fire-detection.” https://github.com/ [2] J. Gubbi, R.

Buyya, S. Marusic, and M. Palaniswami, “Internet of things arpit-jadon/FireNet-LightWeight-Network-for-Fire-Detection.git, 2019.

(iot): A vision, architectural elements, and future directions,” Future [26] N. Srivastava, G. Hinton, A.

Krizhevsky, I. Sutskever, and R. Salakhut- generation computer systems, vol. 29, no. 7, pp. 1645–1660, 2013. dinov, “Dropout: a simple way to prevent neural networks from over- [3] “Raspberry Pi 3 Model B.” https://www.raspberrypi.org/products/ fitting,” The Journal of Machine Learning Research, vol. 15, no. 1, raspberry-pi-3-model-b/.

Accessed: 2019-14-03. pp. 1929–1958, 2014. [4] P. Foggia, A. Saggese, and M.

Vento, “Real-time fire detection for video- [27] “Twilio.” https://www.twilio.com/. Accessed: 2019-01-05. surveillance applications using a combination of experts based on color, [28] “Amazon Simple Storage Service (Amazon S3).” https://aws.amazon. shape, and motion,” IEEE TRANSACTIONS on circuits and systems for com/s3/. Accessed: 2019-01-05. video technology, vol. 25, no. 9, pp. 1545–1556, 2015.

A. Architecture

TABLE I

Y. Chino, L. P.

A. Architecture

TABLE I

Y. Chino, L. P.