Content uploaded by Shreyas Ramachandran Srinivasan
Author content
All content in this area was uploaded by Shreyas Ramachandran Srinivasan on Dec 30, 2018
Content may be subject to copyright.
A preview of the PDF is not available
Development Of Autonomous Downscaled Model Car Using Neural Networks And Machine Learning
Abstract
In our upcoming world, the number of accidents occurring has increased drastically during the recent years leading to increase in fatal deaths. This is mostly caused by the distractions a driver encounters , for example, texting and driving,less attention span of driver,etc. Due to the above reasons, autonomous cars would be a better option which takes the errors of a driver away from the equation. The proposed concept in the paper is to make an autonomous downscaled model car using a generic RC car as base. We aim to achieve the above by using image processing which is trained by using neural networks to create a model through which autonomous cars are achieved. The hardware components used in this project are Raspberry PI 3 B microcomputer , camera module, HCSR04 ultrasonic sensor. We achieve the following features in our model, (a)Lane detection, (b)Traffic signal identification, (c)Road signs identification, (d)Obstacle detection avoidance, (e)Pedestrian Detection. The user can interface through an application that runs on a Raspberry Pi 3 Model B microcomputer that can be accessed on another computer via a graphical desktop sharing system termed as Remote Frame Buffer (RFB).
... In the study of Truong-Dong Do et al. [8], they achieved road tracking and traffic sign recognition on a custom track with an AI algorithm. In the study of Karni et al. [9], they achieved road tracking and obstacle recognition on a custom track with an AI algorithm. ...
... Regarding the selection of sensors, in the study by Karni et al. [9], Lee and Lam [2], they used a camera and ultrasonic. In their studies, ultrasonic only enabled obstacle detection, and the intelligent driving platform can only stop and cannot avoid obstacles. ...
... However, single-task autonomous driving is not truly autonomous driving. In the study of Karni et al. [9], they trained a CNN model for road tracking and obstacle detection. In the study of Truong-Dong Do et al. [8], they trained a CNN model that achieved road tracking and traffic sign recognition. ...
Road tracking, traffic sign recognition, obstacle avoidance, and real-time acceleration and deceleration are some critical sub-tasks in autonomous driving. This research proposed to use a single-sensor (camera) based intelligent driving platform to achieve multi-task (four subtasks) autonomous driving. We adjusted the function combinations and hyperparameters of the model to improve the model training and model testing performance. The experiments showed that the existing function combinations could not significantly improve the autonomous driving performance, and the loss function had a significant impact on the autonomous driving performance of the model. Therefore, we designed a novel loss function (MFPE) based on multi-task autonomous driving. The models with the MFPE loss function outperformed the original and existing models in model training and actual multi-task autonomous driving performance. Meanwhile, the model with the MFPE loss function achieved multi-task autonomous driving under different lighting conditions, untrained routes, and different static obstacles, which indicates that the MFPE loss function enhances the robustness of the model. In addition, the speed of the intelligent driving platform can reach up to 5.4 Km/h.
... Techniques for developing advanced visual algorithms for rapid object detection are outlined in the following papers. According to a remote-controlled car, Karni et al. [10] constructed a small-sized self-driving vehicle using a vision-based CNN network to mimic self-driving control. Wei et al. [11] combined millimeter-wave radar and vision fusion to detect the obstacle precisely. ...
Our previous work introduced the LW-YOLOv4-tiny and the LW-ResNet18 models by replacing traditional convolution with the Ghost Conv to achieve rapid object detection and steering angle prediction, respectively. However, the entire object detection and steering angle prediction process has encountered a speed limit problem. Therefore, this study aims to significantly speed up the object detection and the steering angle prediction simultaneously. This paper proposes the GhostBottleneck approach to speed the frame rate of feature extraction and add the SElayer method to maintain the existing precision of object detection, which constructs an enhanced object detection model abbreviated as LWGSE-YOLOv4-tiny. In addition, this paper also conducted depthwise separable convolution to simplify the Ghost Conv as depthwise separable and ghost convolution, which constructs an improved steering angle prediction model abbreviated as LWDSG-ResNet18 that can considerably speed up the prediction and slightly increase image recognition accuracy. Compared with our previous work, the proposed approach shows that the GhostBottleneck module can significantly boost the frame rate of feature extraction by 9.98%, and SElayer can upgrade the precision of object detection slightly by 0.41%. Moreover, depthwise separable and ghost convolution can considerably boost prediction speed by 20.55% and increase image recognition accuracy by 2.05%.
... Out of the seven studies, two ( [21] and [22]) predicted steering directions with high accuracy between 86% and 95%. An additional three studies calculated angles using Nvidia and Raspberry Pi devices, each utilizing at least 8GB of GPU for additional power. ...
1. Motivation: Lane detection is a critical component of autonomous vehicles, and essential for ensuring their safe and efficient navigation. Effective navigation solutions for affordable devices could contribute to the increasing use of robots and autonomous vehicles in a wide range of applications.
2. Problem statement: In recent years, several models have been proposed to improve the accuracy and performance of this activity, however, little is known about practical real-time implementations of these models in small-scale vehicles and low-cost devices, with limited computing and memory resources. Additionally, the suitability of deep learning models (DL) for this type of device remains unclear.
3. Approach: This report presents the results of a systematic mapping study aimed at gaining a deeper understanding of the models, techniques, hardware, and systems that have been employed for lane recognition on these devices. Additionally, an experimental evaluation was conducted for testing different models for lane detection in real-time on miniature cars, identifying the best-performing models. Three methods were analysed: (1) a model using traditional image processing techniques, (2) an end-to-end DL model proposed by Nvidia, which directly predicts angles, and (3) an ultrafast lane detection DL model, which predicts lane boundaries. Our hypothesis is that (3) is the best-performing model due to its loss calculation formula, which is based on local and global features and could detect lanes better.
To train the DL models, labelled data is required. In this study, an automated approach was employed for estimating the labels, which represent the steering angles or define the road markings used in the recognition process.
4. Results: The mapping study shows that a few models have been implemented in real-time applications for this kind of device. Among the most criticized aspects of them is the high processing and decision-making times. Another relevant point is that, although the publications provide insights into their approaches consisting of software-hardware systems, they do not allow a comparative evaluation among the models. The experiments conducted in this study enabled a robust and impartial comparative analysis among the methods. The results indicate that Model 1, which utilizes traditional image processing techniques, achieved a frame rate of 758 frames per second (FPS) without using any GPU, and with a memory usage of only 0.22 GB. However, it also demonstrated some scalability issues and lower accuracy compared to the other models, with a correct angle prediction rate of 86%. Among the DL methods, the ultra-fast model (3), which predicts lane boundaries, outperformed the end-to-end Nvidia model (2), which directly predicts steering angles. The former achieved a higher level of performance compared to the latter, with an accuracy of 95% compared to 92%. Despite its superior performance, the ultra-fast model has the disadvantage of requiring a longer processing time when run without a GPU, resulting in a lower frame rate of only 0.18 FPS. However, some adjustments were made to this model, resulting in an increase in its frame rate to 13 FPS while maintaining the same level of accuracy. On the other hand, the end-to-end model maybe simpler to implement and faster but is more prone to producing unreliable results when faced with unbalanced data.
5. Conclusions: In summary, the results of this study demonstrate the adequacy of DL models for addressing the lane detection recognition, even on low-cost devices, as long as minimum capacity and memory requirements are met. These models yield significantly more robust results and offer superior performance in handling variations in geometry and luminosity. A comparative analysis based on accuracy, memory consumption, and processing time is provided and serves as a useful guide for selecting the most appropriate model. It is essential to evaluate other traditional techniques and deep learning models. Finally, further research is needed to improve solutions, such as lighter backbones, libraries and tools, specifically tailored for mobile, embedded, and edge devices.
Purpose
The purpose of this research is to achieve multi-task autonomous driving by adjusting the network architecture of the model. Meanwhile, after achieving multi-task autonomous driving, the authors found that the trained neural network model performs poorly in untrained scenarios. Therefore, the authors proposed to improve the transfer efficiency of the model for new scenarios through transfer learning.
Design/methodology/approach
First, the authors achieved multi-task autonomous driving by training a model combining convolutional neural network and different structured long short-term memory (LSTM) layers. Second, the authors achieved fast transfer of neural network models in new scenarios by cross-model transfer learning. Finally, the authors combined data collection and data labeling to improve the efficiency of deep learning. Furthermore, the authors verified that the model has good robustness through light and shadow test.
Findings
This research achieved road tracking, real-time acceleration–deceleration, obstacle avoidance and left/right sign recognition. The model proposed by the authors (UniBiCLSTM) outperforms the existing models tested with model cars in terms of autonomous driving performance. Furthermore, the CMTL-UniBiCL-RL model trained by the authors through cross-model transfer learning improves the efficiency of model adaptation to new scenarios. Meanwhile, this research proposed an automatic data annotation method, which can save 1/4 of the time for deep learning.
Originality/value
This research provided novel solutions in the achievement of multi-task autonomous driving and neural network model scenario for transfer learning. The experiment was achieved on a single camera with an embedded chip and a scale model car, which is expected to simplify the hardware for autonomous driving.
Self-driving cars are a rapidly advancing technology that is poised to revolutionize the way people travel. The purpose of the self-driving car project is to create a vehicle that can operate autonomously, without the need for a human driver. This technology is being developed to increase safety on the road, reduce traffic congestion, and provide a more efficient and convenient mode of transportation. The project involves a combination of hardware and software development, including sensors, cameras, and machine learning algorithms. Self-driving cars have the potential to reduce the number of accidents caused by human error, improve traffic flow, and provide greater mobility for individuals who are unable to drive. While there are still many technical and regulatory challenges to overcome, self-driving cars represent an exciting future for the automotive industry and society as a whole
A vision-based autonomous driving system can usually fuse information about object detection and steering angle prediction for safe self-driving through real-time recognition of the environment around the car. If an autonomous driving system cannot respond fast to driving control appropriately, it will cause high-risk problems with regard to severe car accidents from self-driving. Therefore, this study introduced GhostConv to the YOLOv4-tiny model for rapid object detection, denoted LW-YOLOv4-tiny, and the ResNet18 model for rapid steering angle prediction LW-ResNet18. As per the results, LW-YOLOv4-tiny can achieve the highest execution speed by frames per second, 56.1, and LW-ResNet18 can obtain the lowest prediction loss by mean-square error, 0.0683. Compared with other integrations, the proposed approach can achieve the best performance indicator, 2.4658, showing the fastest response to driving control in self-driving.
Recent advancements in the field of Machine Learning have sprouted a renewed interest in the area of autonomous cars. Companies use different techniques to develop autonomous cars, including buying several vehicles to develop Advanced Driver Assistance Systems (ADAS), while others use car simulators. Simulators for autonomous cars include a wide variety of different sensors. Some simulators come free or even open-source, while others require to pay for the simulator itself, server time, or each sensor. The quality and focus of each of these vary, with some having LIDAR scanned roads for highly realistic roads, while others have entirely natural and realistic vehicle dynamics.This paper gives an overview of available simulators for supervised and reinforcement learning as well as their use cases.KeywordsAutonomous carsMachine learningSimulation
Deep neural network-based classifiers are known to be vulnerable to adversarial examples that can fool them into misclassifying their input through the addition of small-magnitude perturbations. However, recent studies have demonstrated that such adversarial examples are not very effective in the physical world--they either completely fail to cause misclassification or only work in restricted cases where a relatively complex image is perturbed and printed on paper. In this paper we propose a new attack algorithm--Robust Physical Perturbations (RP2)-- that generates perturbations by taking images under different conditions into account. Our algorithm can create spatially-constrained perturbations that mimic vandalism or art to reduce the likelihood of detection by a casual observer. We show that adversarial examples generated by RP2 achieve high success rates under various conditions for real road sign recognition by using an evaluation methodology that captures physical world conditions. We physically realized and evaluated two attacks, one that causes a Stop sign to be misclassified as a Speed Limit sign in 100% of the testing conditions, and one that causes a Right Turn sign to be misclassified as either a Stop or Added Lane sign in 100% of the testing conditions.
In the modern era, the vehicles are focused to be automated to give human driver relaxed driving. In the field of automobile various aspects have been considered which makes a vehicle automated. Google, the biggest network has started working on the self-driving cars since 2010 and still developing new changes to give a whole new level to the automated vehicles. In this paper we have focused on two applications of an automated car, one in which two vehicles have same destination and one knows the route, where other don't. The following vehicle will follow the target (i.e. Front) vehicle automatically. The other application is automated driving during the heavy traffic jam, hence relaxing driver from continuously pushing brake, accelerator or clutch. The idea described in this paper has been taken from the Google car, defining the one aspect here under consideration is making the destination dynamic. This can be done by a vehicle automatically following the destination of another vehicle. Since taking intelligent decisions in the traffic is also an issue for the automated vehicle so this aspect has been also under consideration in this paper.
Visual understanding of complex urban street scenes is an enabling factor for a wide range of applications. Object detection has benefited enormously from large-scale datasets, especially in the context of deep learning. For semantic urban scene understanding, however, no current dataset adequately captures the complexity of real-world urban scenes. To address this, we introduce Cityscapes, a benchmark suite and large-scale dataset to train and test approaches for pixel-level and instance-level semantic labeling. Cityscapes is comprised of a large, diverse set of stereo video sequences recorded in streets from 50 different cities. 5000 of these images have high quality pixel-level annotations; 20000 additional images have coarse annotations to enable methods that leverage large volumes of weakly-labeled data. Crucially, our effort exceeds previous attempts in terms of dataset size, annotation richness, scene variability, and complexity. Our accompanying empirical study provides an in-depth analysis of the dataset characteristics, as well as a performance evaluation of several state-of-the-art approaches based on our benchmark.
The manual tuning of PI and PID controllers and the restriction in choosing the controller parameters are their major drawbacks. In many simulation studies, the artificial intelligence controllers such as Fuzzy logic, ANN and ANFIS are implemented and compared with the performance of conventional PI controllers. It has to be noted that imprecision and approximate reasoning of fuzzy logic, slow learning rate of ANN and ANFIS are found to be the major drawbacks of these controllers. In this paper fuzzy extreme learning machine algorithm is introduced as a novel technique in DC motor speed control application. The speed and current are processed using fuzzy controllers. The fuzzified outputs of these controllers are the inputs to the FELM controller. The trained FELM controller processes these inputs and produces the switching vector control signal by a fuzzy decision matrix. The Mamdani fuzzy model is used in FELM controller. The performance error plot shows that the error is narrowed down to 1e-5. The extreme machine learning algorithm fastens the learning ability of FELM controller. The FELM controller is introduced in the armature circuit. The performance of four quadrant chopper controlled dc motor is tested for reference speeds below and above rated speeds. The results show that the performance of the proposed FELM controller is very efficient in forward and reverse motoring operations compared to the conventional PI controller.
In this paper, convolutional neural network models were developed to perform plant disease detection and diagnosis using simple leaves images of healthy and diseased plants, through deep learning methodologies.Training of the models was performed with the use of an open database of 87,848 images, containing 25 different plants in a set of 58 distinct classes of [plant, disease] combinations, including healthy plants. Several model architectures were trained, with the best performance reaching a 99.53% success rate in identifying the corresponding [plant, disease] combination (or healthy plant). The significantly high success rate makes the model a very useful advisory or early warning tool, and an approach that could be further expanded to support an integrated plant disease identification system to operate in real cultivation conditions.
We introduce a variational Bayesian neural network where the parameters are governed via a probability distribution on random matrices. Specifically, we employ a matrix variate Gaussian \cite{gupta1999matrix} parameter posterior distribution where we explicitly model the covariance among the input and output dimensions of each layer. Furthermore, with approximate covariance matrices we can achieve a more efficient way to represent those correlations that is also cheaper than fully factorized parameter posteriors. We further show that with the "local reprarametrization trick" \cite{kingma2015variational} on this posterior distribution we arrive at a Gaussian Process \cite{rasmussen2006gaussian} interpretation of the hidden units in each layer and we, similarly with \cite{gal2015dropout}, provide connections with deep Gaussian processes. We continue in taking advantage of this duality and incorporate "pseudo-data" \cite{snelson2005sparse} in our model, which in turn allows for more efficient sampling while maintaining the properties of the original model. The validity of the proposed approach is verified through extensive experiments.