Tire Wear Detection for Accident Avoidance Employing Convolutional Neural Networks Tire Wear Detection for Accident Avoidance Employing Convolutional Neural Networks S M Mynul Karim Dept of Computer S[.]
2021 8th NAFOSTED Conference on Information and Computer Science (NICS) Tire Wear Detection for Accident Avoidance Employing Convolutional Neural Networks S.M Mynul Karim Yeaminur Rahman Md Abdul Hai Rezwana Mahfuza Dept of Computer Science and Engineering Brac University Dhaka, Bangladesh s.m.mynul.karim@ g.bracu.ac.bd Dept of Computer Science and Engineering Brac University Dhaka, Bangladesh yeaminur.rahman@ g.bracu.ac.bd Dept of Computer Science and Engineering Brac University Dhaka, Bangladesh md.abdul.hai@ g.bracu.ac.bd Dept of Computer Science and Engineering Brac University Dhaka, Bangladesh rezwana.mahfuza@ g.bracu.ac.bd Abstract—Tires are one of the most essential components of a vehicle, as they actively contribute to driving dynamics However, they are often among the most overlooked when it comes to proper scrutiny and maintenance More often than not, the general masses are found to be negligent of the condition of their tires Treadwear and sidewall damage occur in abundance, and not tending to these problems can have devastating consequences in the long run There is an innumerable number of road accident cases reported which were found to have been caused due to use of damaged and worn-out tires, and these occurrences are more prevalent in highways and during the rainy season Despite being a widespread issue, many people are unable to identify good usable tires from worn-out ones, increasing their likelihood of using dangerous unsafe tires on roads This paper introduces a model that can differentiate between good and worn-out tires, which has been implemented using Image Processing The model takes external pictures of tires provided by the user as input and provides a verdict on its condition after comparing them with the model’s dataset using the machine learning algorithms DenseNet and MobileNet This model has been made keeping in mind that it can be further used with appropriate hardware for implementing in real-life applications By enforcing said implementation by the concerned regulatory bodies, tire-related accidents can be sharply reduced and damage to human life and property can be prevented on public roads Index Terms—Tires, Treadwear, Image Processing, CNN, DenseNet , MobileNet , Accident Avoidance I I NTRODUCTION Tires are one of the oldest inventions of mankind, which have evolved to suit the requirements of the period At present nearly all road-going vehicles are reliant on the use of tires Road tires for vehicles like cars, motorbikes, cycles, and trucks are usually made of a combination of around 200 raw materials [1] Beads and belts of tires are mostly made from steel, while the tread, sidewall, and filling are composed of rubber [2] Rubber is an excellent material for tires, as strong, remain mostly unaffected by temperature change and create less friction with the road surface while providing adequate traction [3] However, the demerit of using rubber is that it is not as durable and is prone to damage over time Tires with smoother tread provide maximum grip on dry tarmac, hence they are used in motorsports On the contrary, they fail to maintain traction on uneven and moist surfaces 978-1-6654-1001-4/21/$31.00 ©2021 IEEE In order to maintain grip in all weather conditions, tires have grooves in their tread which allow the water to escape when roads are wet These grooves also enable the tire to flex for maximizing surface contact with the road, helping to provide more traction As mentioned earlier, the rubber in the tire tread gets wornout over time Treadwear can be of several types namely center wear, side wear, cupping, feathering, etc Unwarranted treadwear can lead a tire to bald, causing the tire to bear several negative effects like excessive heat buildup, increased risk of hydroplaning, loss of air pressure, and difficulty of handling in snow [4] These conditions can result in various consequences such as difficulty in stopping and decreased traction or at worst could cause tire blowout or trigger a fire [5] According to the National Highway Traffic Safety Administration USA (NHTSA), bad tires are responsible for around 35% of crashes caused due to the vehicle itself [6] The numbers look worse for trucks, as an estimated 223 persons have died in large-truck crashes because of tires between 2009 and 2013 [7] A study from 2019 shows that tire-related death crashes far exceed deaths caused by cell phone distraction [8] All these are indicative of the fact that worn-out tires are not in any manner suitable for vehicular commute, especially on highways The dangers of using worn-out tires are not common knowledge among many Some are not informed on the matter, while others remain ignorant about the consequences One in three Americans fails to identify if their tire is bald [9] Tires are designed in a way that they need to be replaced after withstanding a certain number of miles The NHTSA recommends changing tires after the tread depth is 2/32” [10] Despite the presence of these regulations, many people leave their tires unchanged just to save a quick buck without considering the long-term implications The objectives of this paper are mentioned as follows: • Prepare a model to differentiate between healthy and worn-out tires • Train the model with an appropriate dataset consisting of images of tires in different conditions using image processing 364 2021 8th NAFOSTED Conference on Information and Computer Science (NICS) Compare the machine learning algorithms DenseNet121, DenseNet201, and MobileNet to determine the most suitable algorithm for detecting worn-out tires Whether it is due to being clueless or ignorant, it can be stated that people are not conscious enough about their tire health By implementing this model, road accidents caused due to faulty tires can be reduced by a large margin Furthermore, this technology may be employed discretely as well as scaled for mass application • II L ITERATURE R EVIEW K B Sing et al provided an overview of the use of machine learning and deep learning approaches to tires [11] The authors highlighted the progress of machine learning in the tire business, as well as many of its practical applications, in the research Furthermore, the study demonstrated various unique use cases by using ML approaches, such as determining tire properties such as stiffness and grip by implementing Linear Regression, Random Forest, Gradient Boosting, and XG Boosting algorithms Moreover, the authors highly recommended the CNN algorithm’s VGG 16 architecture for tire survey automation To predict the aquaplaning performance of tires, T Weyde et al implemented different Machine Learning algorithms [12] Using Standard Neural Networks (SNN) the authors developed a set of features responsible for hydroplaning of the tread grooves and applied the algorithm on their original dataset Through the study, the team was able to show that ML-based models had improved prediction performance of the used heuristic formula in a cross-validated test over the conventional methods like FT baseline To assess tire state, J E Siegel [13] established a densely connected convolutional neural network based on smartphone images The authors constructed VGG-16 and DenseNets, with the DenseNet outperforming them all Afterwards, they tweaked the hyperparameters of DenseNet containing four blocks of 15 layers more than 10 times for seeing validation loss and average precision They subsequently explored different filters and activation functions on the baseline pictures of the first layer to see whether they could tell the difference between cracked and regular tires Furthermore, they created an efficient CNN classifier for DenseNet, allowing drivers to snap a photo using their phones and assess the tire status via an HTTP portal But colors, shadows, angles of light exposure, and contrast-based picture segmentation should be precisely taken into account to get superior results Additionally, when collecting with a mobile device, redundant data may get added, resulting in a negative outcome that fails to highlight key regions W Kazmi et al [14] proposed a system that identifies a comprehensive technique for detecting and recognizing tire codes on tire sidewalls The authors detected tire circularity using Circular Hough Transform which unwrapped the tires into a rectangular patch and uses CNN architecture for text recognition The research also proposed a Histogram of Oriented Gradients to localize the tire codes Apart from that, the authors used two separate CNNs to improve the accuracy of text recognition The task of identifying the texts becomes critical on worn tires However, the result achieved by the research could not be compared by related works and there is still enough room for improving the text detector III A LGORITHMS A CNN Architecture and Transfer Learning Models In several real-world applications, convolutional neural networks have demonstrated promising outcomes Yann LeCun initially used the phrase in 1980, and AlexNet popularized it in 2012 It has demonstrated outstanding performance in object recognition, image segmentation, machine translation, and other areas Convolution is the central component of CNN and also a sort of linear function that is used to extract features The kernel, which is made up of a small number of arrays, is in charge of this Convolutions, nonlinearities, pooling, and classification are the basic operations of a Convolutional Neural Network [15] [16] [17] Fig CNN architecture Fig illustrates the general working model of the CNN architecture The tire in the figure represents the input images of the proposed dataset of tires In the diagram, it is visible how the data is processed in separate phases of convolutional layers, and how different characteristics are extracted for different purposes • DenseNet: DenseNet is a network design that focuses on enhancing deep learning networks while also making them more efficient to train by utilizing shorter connections between layers Each layer in the network is connected to all other layers deeper in the network For example, the first layer is connected to the second, third, fourth, and so on Among different types of Dense Convolutional Network models, DenseNET121 and DenseNET201 are very popular • MobileNet: The MobileNet model is a connectivity method primarily introduced in [18] that incorporates convolution as its fundamental unit It consists of two depthwise separated layers, one for filtering and the other for integrating, leading to substantial reductions in processing and framework complexity Excluding the ultimate fully connected layer, each layer is represented by a batch normalization 365 2021 8th NAFOSTED Conference on Information and Computer Science (NICS) for gradient descent to improve convergence and a rectified linear activation function [19] Later on, two basic global hyper-parameters are proposed with the model to conveniently negotiate on delay and precision with fewer convolution kernels [20] Additionally, it enables the predictive controller to select the most suitable sized framework for the application depending on the issue requirements of users some of the photos were hard to recognize, the contrast and sharpness of the images were adjusted by 20% and 10%, respectively, to achieve higher precision Afterward, the pixel values were normalized and downscaled to 200*200 pixels in RGB When the preprocessing was done, the accuracy of DenseNET121, DenseNET201, and MobileNet was evaluated to identify the most efficient method for the proposed system to identify worn-out tires at highway checkpoints V DATASET IV W ORKFLOW OVERVIEW To ensure a desirable outcome, thorough preparation and stage process implementation is essential Fig depicts the workflow paradigm for the proposed scheme A Dataset Description Due to the unavailability of the dataset of tires, the data had to be collected by self through the team’s endeavor The dataset was acquired using different sources including internet scraping and self-sourcing As a result, all the data contains realistic images, and that will help to train the model better The dataset contains binary features with 228 images The features were labeled as Good tires and Bald/ worn-out tires Fig Sample Images From Collected Dataset In above, Fig represents the sample real-life images from the collected dataset of tires The tire on the left is in fine shape and ready to proceed on the street, whereas the tire on the right is worn-out and dangerous, simulating countless injuries and deaths from road accidents B Dataset Pre-processing Fig Proposed System Architecture As appropriate datasets were not obtainable at the initial stage, two techniques had been used to acquire data One of them was internet scraping, which is the gathering of data from credible online sources, and the other was self-sourcing, which is the acquisition of essential pictures by ourselves Following that, the raw data was cleaned and transformed into a more generic format The processed data was then separated into train, test, and split categories according to necessity Since The dataset was collected from multiple sources, that’s why to give a generalized format and for increasing the efficiency of training all the images were converted to portable document format Besides, some images were trimmed because of invalid formats and not having enough information Afterwards, all the data were split randomly into 80% for Training and 20% for Testing Additionally, a comparative study conducted by Ariateja.D [21] shows that enhancing contrast and denoising images can significantly reduce the Mean Brightness Error and Speckle Noise Strength of an image That is why, to capture all the detail and the patterns of tires, the sharpness of images was increased by 10% Moreover, the image was also tweaked with a 20% contrast enhancement For faster training, all the image 366 2021 8th NAFOSTED Conference on Information and Computer Science (NICS) pixels were rescaled between and to to 255 Furthermore, all the images were converted into 200*200 pixels in RGB format The images were trained with a batch size of 10 and in binary class mode VI C OMPARATIVE S TUDY A ND R ESULT A NALYSIS A Result Obtaining Procedure Every CNN architecture was run individually to achieve the required result All the models were trained with 50 epochs, including a learning rate of 0001 with Adam optimizer The machine used to train the model runs on the Windows 10 operating system, with an intel core i7 - 6500u processor and 8GB of RAM Moreover, the computer has an Nvidia 940MX GPU, which was used to train the model Additionally, it took around hours to train each model independently B Result Analysis To determine the most suitable algorithm to detect worn-out tires it is essential to compare different evaluation measures of the methods for better perception Fig Confusion Matrix of MobileNet TABLE I P ERFORMANCE EVALUATION OF ALGORITHMS USED Accuracy Precision F1-Score Recall DenseNet121 89.13% 89% 89% 89% DenseNet201 91.30% 92% 91% 91% both had an accuracy of 86.96% MobileNet outperforms all other models in this group as well, with perfect accuracy of 100%, whereas DenseNet201 had an accuracy of 95.65% and DenseNet121 had an accuracy of 91.30% MobileNet 95.65% 96% 96% 96% VII C ONCLUSION AND F UTURE W ORKS DenseNet121, DenseNet201, and MobileNet were compared, and a conclusion was reached that MobileNet surpasses all of them with an accuracy of 95.65% The accuracy, precision, F1-score, and recall for DenseNet121, DenseNet201, and MobileNet are analyzed in TABLE I From Fig 4, a confusion matrix is illustrated with actual and predicted outcomes for both bald and good tires with the MobileNet model From the matrix, it is observed that from the actual 23 good tires, 21 tires are predicted as good tires while the remaining two are bald tires, yielding an accuracy of 91.30 % Following that, it was also observed that DenseNet accurately predicted all 23 bald tires from the actual 23 bald tires, with remarkable accuracy of 100 % TABLE II ACCURACY COMPARISON OF USED ALGORITHMS Models DenseNet201 DenseNet121 MobileNet Good Tires 86.96% 86.96% 91.30% Bad Tires 95.65% 91.30% 100% Individual accuracy to identify whether tires are excellent or bald is also computed for DensNet201, DensNet121 According to TABLE II, MobileNet had an accuracy of 91.30 % for detecting good tires, whereas DenseNet201 and DenseNet121 In this paper, the goal was to establish a model by implementing image processing algorithms in order to determine the condition of tire wear, that is if the tire is in a healthy condition or has worn-out After working on the dataset prepared with original data consisting of further enhanced scrubbed images and in real life taken pictures, a working model was initiated, and after testing several machine learning algorithms for image processing, the best algorithm for fulfilling that purpose was found The objectives of the paper were successfully met, as the model for differentiating faulty tires had been implemented effectively, and the best algorithm for it was MobileNet This model had higher accuracy and precision than both of the DenseNet models used in the paper while boasting a 100% accuracy in identifying bad tires in general There were certain limitations of the work executed in this paper, as merely classifying tires as good or bad will not solve the problems completely, and being able to predict other metrics like the remaining average lifespan would have been more beneficial However, the largest hurdle in doing so was the lack of an already existing appropriate dataset The availability of a larger dataset would have helped in obtaining more varieties of data, thus enabling the possibility of classifying the images of worn-out tires by their relative condition into several categories In the future, such classifications can be used to determine remaining tire life and compare the impact of different tread patterns in tires 367 2021 8th NAFOSTED Conference on Information and Computer Science (NICS) By developing an application based on the presented model, regular people can get access to the benefits of monitoring tire safety from the tip of their hands The model could be further co-developed with a hardware-based system built on IoT Such a system can be deployed on highway entry booths by the road safety department to monitor tires of the vehicles entering the highways to remove the risk factor of accidents caused by worn-out tires There are many other vehicle and passenger related issues responsible for accidents on roads that need to be addressed Nevertheless, tackling the issue of bad tires can serve as an easy yet effective first step towards achieving complete genuine road safety [9] [10] [11] [12] [13] R EFERENCES [1] “An unknow object: the tire materials.” https://thetiredigest.michelin.com/an-unknown-object-the-tire-materials [2] A Gursel, E Akca, and N Sen, “A review on devulcanization of waste tire rubber,” Periodicals of Engineering and Natural Sciences, vol 6, pp 154–160, 10 2018 [3] W Panwiriyarat, N Saetung, H Badawy, C Khaokong, P Pasetto, I Campistron, A Nourry, S Pascual, L Fontaine, T Cutrigth, J.-L Mouget, V Tanrattanakul, and J.-F Pilard, “Natural rubber: An old material for new applications,” 10 2012 [4] T F Fwa, S S Kumar, K Anupam, and G P Ong, “Effectiveness of tire-tread patterns in reducing the risk of hydroplaning,” Transp Res Rec., vol 2094, no 1, pp 91–102, 2009 [5] Y Li, S Zuo, L Lei, X Yang, and X Wu, “Analysis of impact factors of tire wear,” Journal of Vibration and Control, vol 18, pp 833–840, 05 2012 [6] “Critical reasons for crashes investigated in the national motor vehicle crash causation survey.” https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/812115 [7] Newlouth, “Bad tires: An underestimated accident factor.” https://www.louthianlaw.com/bad-tires-underestimated-accident-factor/, Feb 2018 [8] BestRide Staff, “Tire-related crash deaths exceed deaths due to cellphone distraction according to safety experts.” https://bestride.com/news/tire- [14] [15] [16] [17] [18] [19] [20] [21] 368 related-crash-deaths-exceed-deaths-due-to-cellphone-distractionaccording-to-safety-experts, Feb 2019 A J Hawkins, “One in three american drivers can’t tell if their tires are bald.” https://www.theverge.com/2018/6/24/17490060/tiretread-bald-safety-nhtsa, June 2018 “Tire-Related factors in the Pre-Crash phase.” https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/811617 K Singh, F Petry, P Alipour, and D Khadraoui, “Application of machine learning & deep learning techniques in the context of use cases relevant for the tire industry,” in Reifen – Fahrwerk – Fahrbahn, pp 241– 268, VDI Verlag, 2019 T Weyde, G Slabaugh, G Fontaine, and C Bederna, “Predicting aquaplaning performance from tyre profile images with machine learning,” in Lecture Notes in Computer Science, pp 133–142, Berlin, Heidelberg: Springer Berlin Heidelberg, 2013 J E Siegel, “Automotive diagnostics as a service: An artificially intelligent mobile application for tire condition assessment,” in Artificial Intelligence and Mobile Services – AIMS 2018 (M Aiello, ed.), (Cham), pp 172–184, Springer International Publishing, 2018 W Kazmi, I Nabney, G Vogiatzis, P Rose, and A Codd, “Vehicle tire (tyre) detection and text recognition using deep learning,” in 2019 IEEE 15th International Conference on Automation Science and Engineering (CASE), pp 1074–1079, IEEE, 2019 P Jansson, “Single-word speech recognition with convolutional neural networks on raw waveforms,”yrkesh ogskolan arcada,2018.” W Ayadi, W Elhamzi, I Charfi, and M Atri, “Deep CNN for brain tumor classification,”neural process,” Lett, 2021 R Yamashita, M Nishio, R K G Do, and K Togashi, “Convolutional neural networks: an overview and application in radiology,” Insights Imaging, vol 9, no 4, pp 611–629, 2018 L SIfre and S Mallat, “Rigid-motion scattering for texture classification,” 2014 eprint: 1403.1687 S Ioffe and C Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” pp 448–456, 2015 eprint: 1502.03167 A Howard, M Zhu, B Chen, D Kalenichenko, W Wang, T Weyand, M Andreetto, and H Adam, “Mobilenets: Efficient convolutional neural networks for mobile vision applications,” 04 2017 eprint: 1704.04861 D Ariateja, I Ardiyanto, and I Soesanti, “A review of contrast enhancement techniques in digital image processing,” in 2018 4th International Conference on Science and Technology (ICST), pp 1–6, 2018 ... comprehensive technique for detecting and recognizing tire codes on tire sidewalls The authors detected tire circularity using Circular Hough Transform which unwrapped the tires into a rectangular... 16 architecture for tire survey automation To predict the aquaplaning performance of tires, T Weyde et al implemented different Machine Learning algorithms [12] Using Standard Neural Networks... monitor tires of the vehicles entering the highways to remove the risk factor of accidents caused by worn-out tires There are many other vehicle and passenger related issues responsible for accidents