Vulnerable road users (VRUs), including pedestrians, cyclists, and motorcyclists, account for approximately 50% of road traffic fatalities globally, as per the World Health Organization. In these scenarios, the accuracy and fairness of perception applications used in autonomous driving become critical to reduce such risks. For machine learning models, performing object classification and detection tasks, the focus has been on improving accuracy and enhancing model performance metrics; however, issues such as biases inherited in models, statistical imbalances and disparities within the datasets are often overlooked. Our research addresses these issues by exploring class imbalances among vulnerable road users by focusing on class distribution analysis, evaluating model performance, and bias impact assessment. Using popular CNN models and Vision Transformers (ViTs) with the nuScenes dataset, our performance evaluation shows detection disparities for underrepresented classes. Compared to related work, we focus on metric-specific and cost-sensitive learning for model optimization and bias mitigation, which includes data augmentation and resampling. Using the proposed mitigation approaches, we see improvement in IoU(%) and NDS(%) metrics from 71.3 to 75.6 and 80.6 to 83.7 for the CNN model. Similarly, for ViT, we observe improvement in IoU and NDS metrics from 74.9 to 79.2 and 83.8 to 87.1. This research contributes to developing reliable models while addressing inclusiveness for minority classes in datasets. Code can be accessed at: BiasDet. ... Read more

System developers and automotive manufacturers have proposed and used advanced driver assistance systems to enable the deployment of next-generation applications in connected vehicles, such as cooperative perception and vehicle-to-everything communication, while addressing autonomy-related challenges. The example deployment of these applications and systems, including proof-of-concept levels, generally depends on vehicle sensor suites, communication units, large-scale memory systems, and high-performance computing (HPC) units, which are together responsible for sensing the vehicle's environment, efficient data processing, and application deployment using machine learning models or rule-based algorithms. These models and algorithms are generally computationally complex as they process sensed and transformed data using statistical algorithms and deep learning models, such as those using convolutional operations and attention mechanisms, for tasks including decision-making, actuation, system analysis, environment sensing, monitoring, and infotainment applications. The computational complexity of these algorithms and models further increases upon scaling, primarily due to high data volumes, which also requires deep convolutions or similar operations for data processing. These operations require high-performance computing units to meet the application's operational and performance requirements, including latency, throughput, and accuracy.

Generally, the AI models used in connected vehicle applications are designed with a prime focus on model performance metrics; however, their high energy consumption and resulting carbon footprints are often overlooked. Recent studies have shown that the computing requirements for autonomous and connected vehicles can themselves become a significant component of overall energy consumption. For example, large-scale deployment of on-board AI across a global vehicle fleet could generate carbon emissions comparable to those of today's entire data center infrastructure. The computing hardware inside autonomous and connected vehicles can consume hundreds to over a thousand watts when running multiple perception and decision models simultaneously. Since these vehicles are battery-powered, this computing energy directly reduces driving range and increases operational cost. At fleet scale, this translates into a substantial carbon footprint, even when vehicles are electric. Therefore, energy efficiency is a primary design requirement, not only for sustainability but also for maintaining vehicle usability, battery longevity, and cost-efficiency. This thesis addresses the disparity between the strong research focus on model accuracy and the limited focus on energy usage by developing and evaluating an energy-aware adaptive framework for AI-driven vehicular services, such as non-safety-critical perception and high-definition mapping applications. The framework achieves energy and runtime improvements through energy-aware training, resource allocation, and adaptive deployment of computationally intensive models.

Previous research has proposed developing energy-efficient solutions in the hardware and software domains. For example, hardware-related energy-efficient solutions include transitioning from high-end graphical processing units to specialized AI accelerators and integrated circuits, which can process neural networks and related operations more efficiently. Similarly, architectural shifts and software-level solutions include transitioning from the centralized computing approach to dedicated edge computing and tiny machine learning solutions. However, the research scope remains within hardware-software co-design and optimization. By specifically targeting software-level optimizations, this thesis explores approximate computing (AxC) as a mechanism to utilize the error resilience of AI models in the perception and latency-tolerant applications of the vehicle-edge computing ecosystem. By balancing a trade-off between quality of experience and energy efficiency, AxC provides opportunities to reduce on-board energy demands and resulting carbon footprints of vehicle and edge devices while maintaining acceptable application performance levels. To explore and optimize the trade-off between model performance and energy consumption for connected autonomous vehicle applications, the following research questions are addressed:

1) What are the requirements for enabling energy efficiency in data-intensive vehicular services?

2) Which components can enable task deployments energy-efficiently and collaboratively in vehicle-edge-cloud computing?

3) How can energy-efficient components be integrated into an energy-aware adaptive software framework?

4) Can the framework effectively balance the trade-off between energy efficiency and performance in vehicle-edge-cloud computing scenarios?

Building upon existing research and knowledge on energy-efficient computing, this thesis addresses the above-mentioned questions. By addressing (RQ1), this research identifies the technical and operational requirements to enable and integrate energy efficiency into data-intensive vehicular services. These functional requirements include performing high-level computations with minimal energy use and efficiently processing large data streams on edge devices. Secondly, to design and develop energy-saving components (RQ2), the thesis proposes software-level approximation schemes combined with variational inference for both training-time and post-training model optimization and acceleration. Third, contributing to (RQ2) and (RQ3), the research explores ML model partitioning and computing resource allocation mechanisms to utilize the distributed and heterogeneous nature of the vehicle-edge-cloud environment for distributed training and inference. These explorations aim to meet service-level objective deployment using lookup table-based mechanisms. Addressing (RQ3) and (RQ4), the thesis integrates these components into an energy-aware adaptive software framework. This framework provides optimized model training and deployment strategies for distributed training and inference on heterogeneous computing resources, while effectively balancing the trade-off between energy efficiency and on-device application performance.

This thesis utilizes the design science methodology, adapting principles from the Information Systems Research Framework (design-as-a-search-process). This approach ensures research rigor to develop artefacts based on the application domain and existing theoretical knowledge. Further within the process, it adds design knowledge to the existing knowledge base of the application domain. As the development cycle of the research methodology includes tests and experiments, the effectiveness of the developed artefacts can be seen through test and experimental evaluation. Applying the proposed software approximation schemes, model partitioning, resource allocation, and adaptive deployment strategies on the state-of-the-art models shows up to 40% improvements in energy saving for less than 7% quality or model performance degradation when compared to the full precision and central computing methods. Software approximation schemes include the design of approximate multipliers, probabilistic approximation mechanisms, approximating convolutional, and fully connected layers for CNNs/DNNs. For the next-generation and memory/compute-intensive vision transformer models, this work proposes software-level approximation schemes based on variational inference, combined with post-training quantization and quantization-aware training, which show up to 35% improvements in energy efficiency for 6-8% quality loss. As the backbone of these next-generation vision models also includes multi-precision operands such as 8-bit, 16-bit, and 32-bit in layers and channels, the research also explores the advantage of mixed-precision operation to facilitate a balanced trade-off between models' energy usage and accuracy.

This research is among the first to investigate energy-aware requirements for application deployment beyond the traditional approach that generally focuses on cloud-based offloading mechanisms and model compression in the context of connected vehicle services and systems. The evaluation of the energy-aware framework on the popular edge devices shows the contribution of the thesis within the scope of distributed model computing using edge AI and sustainable computing practices. The research is set within the area of tiny machine learning and green AI principles. Future research can further develop adaptive algorithms that dynamically optimize energy use in real-time and investigate predictive models under varying conditions. Additionally, exploring the integration of Approximate Computing with emerging technologies like neuromorphic computing can improve processing efficiency in vehicular systems. ... Read more

Driving assist applications and connected autonomous vehicle systems are supported using AI models and algorithms, which process and analyze heavy data volumes. High-performance computing units and large memory systems support these models, algorithms, and applications, which results in additional onboard energy consumption. The current trend is also towards full electrification of vehicles and increasing connectivity in the vehicular ecosystem to support collaborative and distributed applications using vehicle-edge-cloud computing. However, with the increased focus on model performance and improving the accuracy of these models and applications, the issue of high-performance computing requirements and resulting energy consumption are overlooked. The problem becomes more challenging and complex for resource-constrained edge devices, which are battery-dependent and have limited memory and computing power. This paper proposes components for an adaptive framework to reduce energy consumption by balancing model accuracy. The contributions include proposing and integrating model partition mechanisms, adaptive deployment across edge devices and approximation strategies for the models. By integrating these components, this framework supports energy-aware development across various platforms. The approach offers a sustainable method for computing and communication-oriented applications within the vehicular ecosystem. ... Read more

Deploying scalable Vision Transformer applications on mobile and edge devices is constrained by limited memory and computational resources. Existing model development and deployment strategies include distributed computing and inference methods such as federated learning, split computing, collaborative inference and edge-cloud offloading mechanisms. While these strategies have deployment advantages, they fail to optimize memory usage and processing efficiency, resulting in increased energy consumption. This paper optimizes energy consumption by introducing adaptive model partitioning mechanisms and dynamic scaling methods for ViTs such as EfficientViT and TinyViT, adjusting model complexity based on the available computational resources and operating conditions. We implement energy-efficient strategies that minimize inter-layer communication for distributed machine learning across edge devices, thereby reducing energy consumption from data flow and computation. Our evaluations on a series of benchmark models show improvements, including up to a 32.6% reduction in latency and 16.6% energy savings, while maintaining mean average precision sacrifices within 2.5 to 4.5% of baseline models. These results show that our proposal is a practical approach for improving edge AI sustainability and efficiency. ... Read more

Vision transformer (ViTs) models have shown higher accuracy, robustness and large volume data processing ability, creating new baselines and references for perception tasks. However, these advantages require large memory and high-performance processors and computing units, which makes model adaptability and deployment challenging within resource-constrained environments such as memory-restricted and battery-powered edge devices. This paper addresses the model deployment challenges by proposing a model approximation approach VI-ViT, for edge deployment using variational inference with mixed precision for processing multi-modalities, such as point clouds and images. Our experimental evaluation on the nuScenes and Waymo datasets show up to 37% and 31% reduction in model parameters and Flops while maintaining a mean average precision of 70.5 compared to 74.8 of the baseline model. This work presents a practical deployment approach for approximating and optimizing Vision Transformers for edge AI applications by balancing model metrics such as parameters, flops, latency, energy consumption, and accuracy, which can easily be adapted to other transformer models and datasets. ... Read more

Recent advancements in hardware and software systems have been driven by the deployment of emerging smart health and mobility applications. These developments have modernized the traditional approaches by replacing conventional computing systems with cyber–physical and intelligent systems combining the Internet of Things (IoT) with Edge Artificial Intelligence. Despite the many advantages and opportunities of these systems within various application domains, the scarcity of energy, extensive computing needs, and limited communication must be considered when orchestrating their deployment. Inducing savings in these directions is central to the Approximate Computing (AxC) paradigm, in which the accuracy of some operations is traded off with energy, latency, and/or communication reductions. Unfortunately, the dynamics of the environments in which AxC-equipped IoT systems operate have been paid little attention. We bridge this gap by surveying adaptive AxC techniques applied to three emerging application domains, namely autonomous driving, smart sensing and wearables, and positioning, paying special attention to hardware acceleration. We discuss the challenges of such applications, how adaptive AxC can aid their deployment, and which savings it can bring based on traits of the data and devices involved. Insights arising thereof may serve as inspiration to researchers, engineers, and students active within the considered domains. ... Read more

In recent years, there has been a notable increase in the size of commonly used image classification models. This growth has empowered models to recognize thousands of diverse object types. However, their computational demands pose significant challenges, especially when deploying them on resource-constrained edge devices. In many use cases where a model is deployed on an edge device, only a small subset of the classes will ever be observed by a given model instance. Our proposed test-time specialization of dynamic neural networks allows these models to become faster at recognizing the classes that are observed frequently, while maintaining the ability to recognize all other classes, albeit slightly less efficient. We benchmark our approach on a real-world edge device, obtaining significant speedups compared to the baseline model without test-time adaptation. ... Read more

Model partitioning is a promising solution to reduce the high computation load and transmission of high-volume data. Within the scope of Edge AI, the fundamentals of model partitioning involve splitting the model for local computing at the edge and offloading heavy computation tasks to the cloud or server. This approach benefits scenarios with limited computing and battery capacity with low latency requirements, such as connected autonomous vehicles. However, while model partitioning offers advantages in reducing the onboard computation, memory requirements and inference time, it also introduces challenges such as increased energy consumption for partitioned computations and overhead for transferring partitioned data/model. In this work, we explore hybrid model partitioning to optimize computational and communication energy consumption. Our results provide an initial analysis of the tradeoff between energy and accuracy, focusing on the energy-aware model partitioning for future Edge AI applications. ... Read more

Autonomous driving services depends on active sensing from modules such as camera, LiDAR, radar, and communication units. Traditionally, these modules process the sensed data on high-performance computing units inside the vehicle, which can deploy intelligent algorithms and AI models. The sensors mentioned above can produce large volumes of data, potentially reaching up to 20 Terabytes. This data size is influenced by factors such as the duration of driving, the data rate, and the sensor specifications. Consequently, this substantial amount of data can lead to significant power consumption on the vehicle. Similarly, a substantial amount of data will be exchanged between infrastructure sensors and vehicles for collaborative vehicle applications or fully connected autonomous vehicles. This communication process generates an additional surge of energy consumption. Although the autonomous vehicle domain has seen advancements in sensory technologies, wireless communication, computing and AI/ML algorithms, the challenge still exists in how to apply and integrate these technology innovations to achieve energy efficiency. This survey reviews and compares the connected vehicular applications, vehicular communications, approximation and Edge AI techniques. The focus is on energy efficiency by covering newly proposed approximation and enabling frameworks. To the best of our knowledge, this survey is the first to review the latest approximate Edge AI frameworks and publicly available datasets in energy-efficient autonomous driving. The insights from this survey can benefit the collaborative driving service development on low-power and memory-constrained systems and the energy optimization of autonomous vehicles. ... Read more

A machine learning model can often produce biased outputs for a familiar group or similar sets of classes during inference over an unknown dataset. The generalization of neural networks have been studied to resolve biases, which has also shown improvement in accuracy and performance metrics, such as precision and recall, and refining the dataset's validation set. Data distribution and instances included in test and validation-set play a significant role in improving the generalization of neural networks. For producing an unbiased AI model, it should not only be trained to achieve high accuracy and minimize false positives. The goal should be to prevent the dominance of one class/feature over the other class/feature while calculating weights. This paper investigates state-of-art object detection/classification on AI models using metrics such as selectivity score and cosine similarity. We focus on perception tasks for vehicular edge scenarios, which generally include collaborative tasks and model updates based on weights. The analysis is performed using cases that include the difference in data diversity, the viewpoint of the input class and combinations. Our results show the potential of using cosine similarity, selectivity score and invariance for measuring the training bias, which sheds light on developing unbiased AI models for future vehicular edge services. ... Read more

Connected vehicular services depend heavily on communication as they frequently transmit data and AI models/weights within the vehicular ecosystem. Energy efficiency in vehicles is crucial to keep up with the fast-growing demand for vehicular data processing and communication. To tackle this rising challenge, we explore approximation and edge AI techniques for achieving energy efficiency for vehicular services. Focusing on data-intensive vehicular services, we present an experimental case study on the high-definition (HD) map using the model partition approach. Our study compares the AI model energy consumption using multiple approximation ratios over embedded edge devices. Based on experimental insights, we further discuss an envisioned approximate Edge AI pipeline for developing and deploying energy-efficient vehicular services. ... Read more

An Autonomous vehicle depends on the combination of latest technology or ADAS safety features such as Adaptive cruise control (ACC), Autonomous Emergency Braking (AEB), Automatic Parking, Blind Spot Monitor, Forward Collision Warning/Avoidance (FCW or FCA). The current trend follows incorporation of these technologies using the Artificial neural network or Deep neural network, as an imitation of the traditionally used algorithms. The concept behind a FCW algorithm is the measure of a distance or warning range which results in an alert to notify or inform the driver regarding the possible collision. The objective of this paper is to propose a collision warning model using the RTMaps framework. The proposed model is based on the forward facing automotive radar providing the sensed input values such as acceleration, velocity, and separation distance to a neural network based classifier algorithm which on the basis of supervised learning alerts the driver of a possible collision. The implementation, precision and accuracy of the classifier and regression algorithm is discussed and compared. The presented model is deployed on the Bluebox 2.0 platform with the RTMaps Embedded framework. ... Read more

Forward Collision Warning (FCW) also referred as the Forward Collision Avoidance (FCA) system has become an essential part of the Autonomous cars or smart vehicles. The FCW System has found its position as one of the foremost automobile safety features included in the present smart vehicles. This paper proposes a Forward collision warning system which includes its subsystem as Forward-looking automotive radar model and a Classifier algorithm developed and designed in the Intempora RTMaps Embedded with NXP Bluebox 2.0. The classifier and regression algorithm work on the principle of supervised learning, the proposed model predicts the collision or no-collision occurrence based on the radar sensed inputs. The implementation, precision and accuracy of the classifier and regression algorithm is presented. The proposed model is deployed on the Bluebox 2.0 platform with the RTMaps Embedded framework. ... Read more