Emergency Braking Distance Prediction in Autonomous Vehicles

Fire detection is a critical task in ensuring the safety of human lives and property. In recent years, deep learning techniques, such as Convolutional Neural Networks (CNNs), have shown promising results in fire detection due to their ability to automatically learn relevant features from images. In this project, we propose a real-time fire detection system using CNNs to detect fires in images or video streams. The proposed system consists of a CNN-based architecture that includes multiple convolutional layers for feature extraction, activation functions for introducing non-linearity, pooling layers for spatial dimension reduction, and fully connected layers for global pattern learning. The architecture is trained on a labelled dataset of fire and non-fire images or videos using optimization algorithms to minimize the loss function. To achieve real-time fire detection, the trained CNN model is integrated into a fire detection system that can process images or video streams in real-time. The system captures images or video frames from a live feed, pre-processes the data, and passes it through the CNN model for fire detection. The system can generate alerts or notifications in real-time when fire is detected, allowing for prompt response and mitigation measures. The performance of the proposed fire detection system is evaluated using metrics such as accuracy, precision, recall, and F1-score. The system is tested on different datasets and under various conditions to assess its accuracy, robustness, and real-time processing capabilities. The results of this project demonstrate the effectiveness of CNNs for real-time fire detection and highlight the potential of the proposed system for enhancing fire safety in various applications, including surveillance systems, smart buildings, and industrial settings. The developed system can be used as an early warning tool for fire detection, enabling timely response and mitigation, and helping to prevent or minimize the damage caused by fires. Keywords: Fire

This paper proposes a new quantitative precipitation estimation (QPE) technique to provide accurate rainfall estimates in complex terrain, where conventional QPE has limitations. The operational radar QPE in Taiwan is mainly based on the simplified relationship between radar reflectivity Z and rain rate R [R(Z) relation] and only utilizes the single-point lowest available echo to estimate rain rates, leading to low accuracy in complex terrain. Here, we conduct QPE using deep learning that extracts features from 3-D radar reflectivities to address the above issues. Convolutional neural networks (CNN) are used to analyze contoured frequency by altitude diagrams (CFADs) to generate the QPE. CNN models are trained on existing rain gauges in northern and eastern Taiwan with the three-year data during 2015–17 and validated and tested using 2018 data. The weights of heavy rains (≧10 mm h-1) are increased in the model loss calculation to handle the unbalanced rainfall data and improve accuracy. Results show that the CNN outperforms the R(Z) relation based on the 2018 rain-gauge data. Furthermore, this research proposes methods to conduct 2-D gridded QPE at every pixel by blending estimates from various trained CNN models. Verification based on independent rain gauges shows that the CNN QPE solves the underestimation of the R(Z) relation in mountainous areas. Case studies are presented to visualize the results, showing that the CNN QPE generates better small-scale rainfall features and more accurate precipitation information. This deep learning QPE technique may be helpful for the disaster prevention of small-scale flash floods in complex terrain.

For model inference of convolutional neural networks (CNNs), we nowadays witness a shift from the Cloud to the Edge. Unfortunately, deploying and inferring large, compute- and memory-intensive CNNs on Internet of Things devices at the Edge is challenging as they typically have limited resources. One approach to address this challenge is to leverage all available resources across multiple edge devices to execute a large CNN by properly partitioning it and running each CNN partition on a separate edge device. However, there currently does not exist a design and programming framework that takes a trained CNN model as input and subsequently allows for efficiently exploring and automatically implementing a range of different CNN partitions on multiple edge devices to facilitate distributed CNN inference. Therefore, in this article, we propose a novel framework that automates the splitting of a CNN model into a set of submodels as well as the code generation needed for the distributed and collaborative execution of these submodels on multiple, possibly heterogeneous, edge devices, while supporting the exploitation of parallelism among and within the edge devices. In addition, since the number of different CNN mapping possibilities on multiple edge devices is vast, our framework also features a multistage and hierarchical design space exploration methodology to efficiently search for (near-)optimal distributed CNN inference implementations. Our experimental results demonstrate that our work allows for rapidly finding and realizing distributed CNN inference implementations with reduced energy consumption and memory usage per edge device, and under certain conditions, with improved system throughput as well.

The convolutional neural networks (CNN), including AlexNet, GoogleNet, VGGNet, etc. extract features for many computer vision problems which are very discriminative. The trained CNN model over one dataset performs reasonably well whereas on another dataset of similar type the hand-designed feature descriptor outperforms the same trained CNN model. The Rectified Linear Unit (ReLU) layer discards some values in order to introduce the non-linearity. In this paper, it is proposed that the discriminative ability of deep image representation using trained model can be improved by Average Biased ReLU (AB-ReLU) at the last few layers. Basically, AB-ReLU improves the discriminative ability in two ways: 1) it exploits some of the discriminative and discarded negative information of ReLU and 2) it also neglects the irrelevant and positive information used in ReLU. The VGGFace model trained in MatConvNet over the VGG-Face dataset is used as the feature descriptor for face retrieval over other face datasets. The proposed approach is tested over six challenging, unconstrained and robust face datasets (PubFig, LFW, PaSC, AR, FERET and ExtYale) and also on a large scale face dataset (PolyUNIR) in retrieval framework. It is observed that the AB-ReLU outperforms the ReLU when used with a pre-trained VGGFace model over the face datasets. The validation error by training the network after replacing all ReLUs with AB-ReLUs is also observed to be favorable over each dataset. The AB-ReLU even outperforms the state-of-the-art activation functions, such as Sigmoid, ReLU, Leaky ReLU and Flexible ReLU over all seven face datasets.

The recent rapid success of deep convolutional neural networks (CNN) on many computer vision tasks largely benefits from the well-annotated Pascal VOC, ImageNet, and MS COCO datasets. However, it is challenging to get ImageNet-like annotations (1000 classes) in the medical imaging domain due to the lack of clinical training in the lay crowdsourcing community. We address this problem by presenting a semi-supervised training method for neural networks with true-class and pseudo-class (un-annotated class) labels on partially annotated training data. The true-class labels are supervised annotations from clinical professionals. The pseudo-class labels are unsupervised clustering of un-annotated data. Our method rests upon the hypothesis of better coherent annotations with discriminative classes leading to better trained CNN models. We validated our method on extra-coronary calcification detection in low dose CT scans. The CNN trained with true-class and 10 pseudo-classes achieved an 85.0% sensitivity at 10 false positives per patient (0.3 false positive per slice), which significantly outperformed the CNN trained with true-class only (sensitivity =56.0% at 10 false positives per patient).

This study presents a data-driven finite element-machine learning surrogate model for predicting the end-to-end full-field stress distribution and stress concentration around an arbitrary-shaped inclusion. This is important because the model’s capacity to handle large datasets, consider variations in size and shape, and accurately replicate stress fields makes it a valuable tool for studying how inclusion characteristics affect material performance. An automatized dataset generation method using finite element simulation is proposed, validated, and used for attaining a dataset with one thousand inclusion shapes motivated by experimental observations and their corresponding spatially-varying stress distributions. A U-Net-based convolutional neural network (CNN) is trained using the dataset, and its performance is evaluated through quantitative and qualitative comparisons. The dataset, consisting of these stress data arrays, is directly fed into the CNN model for training and evaluation. This approach bypasses the need for converting the stress data into image format, allowing for a more direct and efficient input representation for the CNN. The model was evaluated through a series of sensitivity analyses, focusing on the impact of dataset size and model resolution on accuracy and performance. The results demonstrated that increasing the dataset size significantly improved the model’s prediction accuracy, as indicated by the correlation values. Additionally, the investigation into the effect of model resolution revealed that higher resolutions led to better stress field predictions and reduced error. Overall, the surrogate model proved effective in accurately predicting the effective stress concentration in inclusions, showcasing its potential in practical applications requiring stress analysis such as structural engineering, material design, failure analysis, and multi-scale modeling.

We aimed to develop a model to predict visual field (VF) in the central 10 degrees in patients with glaucoma, by training a convolutional neural network (CNN) with optical coherence tomography (OCT) images and adjusting the values with Humphrey Field Analyzer (HFA) 24–2 test. The training dataset included 558 eyes from 312 glaucoma patients and 90 eyes from 46 normal subjects. The testing dataset included 105 eyes from 72 glaucoma patients. All eyes were analyzed by the HFA 10-2 test and OCT; eyes in the testing dataset were additionally analyzed by the HFA 24-2 test. During CNN model training, the total deviation (TD) values of the HFA 10-2 test point were predicted from the combined OCT-measured macular retinal layers’ thicknesses. Then, the predicted TD values were corrected using the TD values of the innermost four points from the HFA 24-2 test. Mean absolute error derived from the CNN models ranged between 9.4 and 9.5 B. These values reduced to 5.5 dB on average, when the data were corrected using the HFA 24-2 test. In conclusion, HFA 10-2 test results can be predicted with a OCT images using a trained CNN model with adjustment using HFA 24-2 test.

When a light beam transmits in free space, it is easily affected by atmospheric turbulence. The effect on transmitted light is equivalent to adding a random noise phase to it, which leads its transmission quality to deteriorate. The method of improving the quality of transmitted beams is usually to compensate for the phase distortion at the receiver by adding reverse turbulence phase, and the premise of this method is to obtain the turbulence phase carried by the distorted beam. The adaptive optics system is the most common way to extract the phase information. However, it is inefficient to be applied to varying turbulence environments due to the fact that a wave-front sensor and complex optical system are usually contained. Deep convolutional neural network (CNN) that can directly capture feature information from images is widely used in computer vision, language processing, optical information processing, etc. Therefore, in this paper proposed is a turbulence phase information extraction scheme based on the CNN, which can quickly and accurately extract the turbulence phase from the intensity patterns affected by atmosphere turbulence. The CNN model in this paper consists of 17 layers, including convolutional layers, pooling layers and deconvolutional layers. The convolutional layers and pooling layers are used to extract the turbulent phase from the feature image, which is the core structure of the network. The function of the deconvolutional layers is to visualize the extracted turbulence information and output the final predicted turbulence phase. After learning a huge number of samples, the loss function value of CNN converges to about 0.02, and the average loss function value on the test set is lower than 0.03. The trained CNN model has a good generalization capability and can directly extract the turbulent phase according to the input light intensity pattern. Using an I5-8500 CPU, the average time to predict the turbulent phase is as low as s under the condition of <inline-formula><tex-math id="M5">\begin{document}$C_{{n}}^2 = 1 \times {10^{ - 14}}\;{{\rm{m}}^{ - 2/3}}$\end{document}</tex-math><alternatives><graphic xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="1-20190982_M5.jpg"/><graphic xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="1-20190982_M5.png"/></alternatives></inline-formula>, <inline-formula><tex-math id="M6">\begin{document}$ 5 \times {10^{ - 14}}\;{{\rm{m}}^{ - 2/3}}$\end{document}</tex-math><alternatives><graphic xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="1-20190982_M6.jpg"/><graphic xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="1-20190982_M6.png"/></alternatives></inline-formula>, and <inline-formula><tex-math id="M7">\begin{document}$1 \times {10^{ - 13}}\;{{\rm{m}}^{ - 2/3}}$\end{document}</tex-math><alternatives><graphic xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="1-20190982_M7.jpg"/><graphic xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="1-20190982_M7.png"/></alternatives></inline-formula>. In addition, the turbulence phase extraction capability of CNN can be further enhanced by improving computing power or optimizing model structure. These results indicate that the CNN-based turbulence phase extraction method can effectively extract the turbulence phase, which has important application value in turbulence compensation, atmospheric turbulence characteristics research and image reconstruction.

Plant disease detection and classification have undergone successful researches using Convolutional Neural Network (CNN); however, due to the intrinsic inability of max pooling layer in CNN, it fails to capture the pose, view and orientation of images.It also requires large training data and fails to learn the spatial relationship of the features in an object.Thus, Capsule Network (CapsNet) is a novel deep learn-ing model proposed to overcome the shortcomings of CNN.We developed an optimized Capsule Network model for classification problem using banana leaf diseases as a case study.The two dataset classes in-clude Bacterial Wilt and Black Sigatoka, with healthy leaves.The developed model adequately classified the banana bacterial wilt, black sigatoka and healthy leaves with a test accuracy of 95%.Its outper-formed three variants of CNN architectures implemented (a trained CNN model from scratch, LeNet5 and ResNet50) with respect to rotation invariance.

A novel application of XAI in squinting models: A position paper

Objective: Forward collision warning (FCW) and automatic emergency braking (AEB) have the potential to prevent or mitigate many large truck crashes. While these systems are known to be effective in passenger vehicles, less is known about their effectiveness in large trucks. The objective of this study was to estimate the effectiveness of these systems in reducing real-world crash rates of large trucks. Methods: Data on Class 8 trucks operating on limited-access highways during 2017–2019 were obtained from SmartDrive Systems. Detailed data on exposure measures and crash circumstances were extracted from video footage by both automated means and manual coding. Crash rates were compared by front crash prevention technology (FCW, AEB, neither), both for all police-reportable crashes overall and for relevant crash types. Results: FCW was associated with a statistically significant 22% reduction in the rate of police-reportable crashes per vehicle miles traveled, and a significant 44% reduction in the rear-end crash rate of large trucks. AEB also was associated with significant reductions—12% overall and 41% for rear-end crashes. Warnings were issued in 31% of rear-end crashes for FCW-equipped trucks. AEB intervened in 43% of rear-end crashes; about two thirds of these interventions involved autobrake activations. On average, speed was reduced by over half between the time of the intervention and impact for both systems. Observed reductions in same-direction sideswipe and roadway departure crashes per mile traveled were smaller in magnitude than those of rear-end crashes; these were consistent with other crash avoidance technologies suspected to be bundled with FCW/AEB in some cases, and very few front crash prevention interventions occurred in these types of crashes. Conclusions: FCW and AEB are effective countermeasures for crashes in which large trucks rear-end other vehicles. Large truck safety is expected to improve as new trucks are increasingly equipped with these systems. FCW has the advantage that some of these systems can be retrofitted to existing trucks, so benefits can be realized sooner and with less investment.

Autonomous vehicles are in full development and vehicles classification is a fundamental part of this new technology. To interact with other objects on the road, vehicles need to be able to identify what is surrounding them. In this paper, we develop a platform named VARC (Vehicle Algorithm for Recognition and Classification) integrating a fully connected neural network to classify different types of vehicles such as trucks. Moreover, VARC considers the detected type to identify the brand of the vehicle using a convolutional neural network. This allows getting valuable characteristics of the vehicle like its color, the number of passengers. By processing pictures taken from security cameras or from ones on vehicles, VARC may help cops identify stolen cars and autonomous vehicles improve their perception of their environment. VARC's neural network is trained on more than 7000 images of cars, trucks, and motorcycles. Results demonstrated the effectiveness of VARC in terms of generating valuable data while minimizing the needed resources.

Objectives/Aim: The present paper deals with building a smart assistant with the aim of assisting the visually impaired people in mobility with confidence by realizing the nearby obstacles and also implement image processing techniques to recognize people. Methods: We use Raspberry Pi 4 which has increased computing performance and is interfaced to the picamera, GPS and GSM modules. Arduino pro mini with buzzer, ultrasonic sensors and vibration motors are used for obstacle detection. Certain libraries important for image processing are used such as OpenCV, Dlib, face detection (Haar cascades, HOG + Linear SVM, or CNNs), espeak for converting text to speech. Programming was implemented through Python and Arduino compiler. Findings: Analysis was carried out using the proposed system for the blind and visually impaired people who could move around comfortably with confidence and were also able to detect objects and recognize people. Hence the proposed system removes the use for the cumbersome white cane in exchange for small and handy modules which can be used in the form of wearables mounted on shoulders, hands and legs to detect the obstacles from multiple directions and provide for a comfortable wear. It was determined that the Raspberry pi 4 was incapable of running CNN detection for which ideally computer GPU was required, so HOG detection method was used instead. Application/Improvements: This project can be implemented mainly in the commercial field of helping visually impaired people with poor eyesight or being completely blind. Industrial applications can be devised and enhanced like robots and machineries along with Security, Identifying and Tracking. Keywords: Face Recognition; Obstacle Detection; Open Source Computer Vision (OpenCV); Deep Learning; Histogram of Oriented Gradients (HOG); Support Vector Machine (SVM); Convolutional Neural Network (CNN)

This paper presents the application of convolutional neural networks (CNNs) for developing a classification system of X-Ray images into three classes: Normal, Benign, and Malignant. For training convolutional neural network models, we use nearly 9000 Vietnamese mammograms, which were collected and annotated by the radiologists of the Vietnam National Cancer Hospital. Our experiments were conducted with the 2 phases with two different data sets. The first phase experimented with multiple training options, as with the weighted loss function, the gray-level pre-processing, and the same number of samples of classes. The results of experiments in the first phase would form the basis for the second phase for developing a mammogram classification system based on the CNN with ResNet 18. The evaluation of system performance using the testing set achieves a macAUC of 0.828247, an average sensitivity of 0.64738, and an average specificity of 0.825670. In particular, the system accuracy achieves an AUC of 0.894291 in predicting the presence of cancer. The evaluation of system performance achieved a macAUC of 0.828247 which is higher than a macAUC of 0.754 for the classifying images into BI-RADS 045, BI-RADS 1, BI-RADS 23 categories when using 7912 breast cancer X-Ray images collected by Vietnam Hanoi Medical University Hospital for ResNet 50 training.

In order to proceed with shipbuilding scheduling involving hundreds of hull blocks of ships, it is important to mark the locations of the hull blocks with the correct block identification number. Incorrect information about the locations and the identification numbers of hull blocks causes disruption in the shipbuilding scheduling process of the shipyard. Therefore, it is necessary to have a system for tracking the locations and identification numbers of hull blocks in order to avoid time loss due to incorrectly identified blocks. This paper proposes a method to mark the identification numbers, which are necessary for the tracking system of hull blocks. In order to do this, 3 CNN (convolutional neural network) models, VGG-19, Resnet-152V2, and Densenet-201, are used to classify the hull blocks. A set of multi-view images acquired from 3D CAD data are used as training data to obtain a trained CNN model, and images from 3D printer-printed hull block models are used for the test of the trained CNN model. The datasets used for training and prediction are Non-Thr and Thr datasets, that each included both binarized and non-binarized datasets. As a result of end-to-end classification experiments with Non-Thr datasets, the highest prediction accuracy was 0.68 with Densenet-201. A total of 4050 experimental conditions were constructed by combining the threadhold of the Thr training and testing dataset. As a result of experiments with Thr datasets, the highest prediction accuracy of 0.96 was acquired with Resnet-152V2, which was trained with a threshold of 72 and predicted with a threshold of 50. In conclusion, the classification of ship hull blocks using a CNN model with binarized datasets of 3D CAD data is more effective than that using a CNN model with non-binarized datasets.