PaperlayerView exact revision on arXiv

Source-linked AI summary

A Systematic Literature Review about the impact of Artificial Intelligence on Autonomous Vehicle Safety

A. M. Nascimento, L. F. Vismari, C. B. S. T. Molina, P. S. Cugnasca, J. B. Camargo Jr., J. R. de Almeida Jr., R. Inam, E. Fersman, M. V. Marquezini, A. Y. Hata

arXiv:1904.02697v1cs.CYcs.AI

TL;DR

Researchers disagree on whether AI enhances AV safety or can jeopardize it, despite the safety-critical stakes of AV operation. This paper reviews the literature and maps the studies to clarify the field. It concludes that future research needs a serious safety agenda, with the reviewed field not yet strongly shaped by safety engineering culture.

  • Problem

    Researchers disagree on whether AI enhances AV safety or potentially jeopardizes it, even though AVs are safety-critical systems.

    Source: §I · p. 1

  • Method

    The paper conducts a systematic literature review to map and organize research on AI and AV safety.

    Source: §I · p. 1

  • Results

    The reviewed literature suggests that AI and AV research is not yet heavily influenced by safety engineering culture, with relatively few studies on fault forecasting, ethics and policies, and dependability and trust.

    Source: §V · p. 12

  • Takeaways & Limitations

    The authors reinforce the need for a serious safety agenda in future studies of AI-based AV systems.

    Source: §V · p. 12

    Some databases did not support searches restricted to title, abstract, and keyword fields, inflating the number of papers retrieved.

    Source: B. Identification of Search String and… · p. 2

Abstract

from arXiv · show

Autonomous Vehicles (AV) are expected to bring considerable benefits to society, such as traffic optimization and accidents reduction. They rely heavily on advances in many Artificial Intelligence (AI) approaches and techniques. However, while some researchers in this field believe AI is the core element to enhance safety, others believe AI imposes new challenges to assure the safety of these new AI-based systems and applications. In this non-convergent context, this paper presents a systematic literature review to paint a clear picture of the state of the art of the literature in AI on AV safety. Based on an initial sample of 4870 retrieved papers, 59 studies were selected as the result of the selection criteria detailed in the paper. The shortlisted studies were then mapped into six categories to answer the proposed research questions. An AV system model was proposed and applied to orient the discussions about the SLR findings. As a main result, we have reinforced our preliminary observation about the necessity of considering a serious safety agenda for the future studies on AI-based AV systems.

I. INTRODUCTION

AI is central to AV operation, but researchers disagree on whether it improves or threatens safety; this review maps both perspectives and describes its research protocol. The protocol sets research questions and searches literature databases using selected terms.

I. INTRODUCTION

  • AVs rely on AI to interpret their environment and make driving decisions, while safety is essential because unsafe operation can threaten people and infrastructure.The review frames disagreement over AI’s safety impact as its central motivation.Source: §I · p. 1
  • The review presents itself as the first study, to the authors’ knowledge, to organize literature covering both perspectives on AI and AV safety.It aims to provide a clear picture of the field and support future studies.Source: §I · p. 1

II. RESEARCH METHODOLOGY

  • The authors use a systematic literature review protocol based on prior guidance for defining questions, searching, selecting studies, mapping data, and reporting findings.The review method is presented as a way to understand the state of the art and identify research directions.Source: §II · p. 1

A. Definition of Research Questions

B. Identification of Search String and Source Selection

  • The search covered seven literature databases without a date limit and used selected synonyms for safety, AI, and autonomous vehicles.Searches were adapted to database syntax and targeted title, abstract, and keyword fields where supported.Source: B. Identification of Search String and… · p. 2

C. Study Selection Criteria and Papers Review

Screening reduced the initial search results to 59 papers, which the authors mapped into six categories aligned with the research questions. The review describes publication patterns, including recent growth and the distribution of papers across conferences and journals.

C. Study Selection Criteria and Papers Review

  • 59 papers were selected from an initial 4,870; the reduction reflected terminology misuse, irrelevant examples, and databases that could not restrict searches to title, abstract, and keywords.The process first retained 230 papers after checking those fields, then 97 after duplicate removal, before later exclusions.Source: C. Study Selection Criteria and Papers… · pp. 2–3

D. Data Mapping

  • The 59 reviewed papers were categorized into six groups corresponding to the six research questions: impact, topics, techniques, problems, findings, and future studies.Source: D. Data Mapping · p. 3

III. DATA ANALYSIS

  • The oldest included study dates to 1987, no study was found from 1991 to 2002 under the search criteria, and 86% of papers were published from 2012 to 2018.The authors report continuous growth in publications since 2014; 2018 was excluded from the decade plot.Source: §III · p. 3
  • 45 papers (76%) were conference publications and 14 (24%) were journal papers; keyword totals exceed 59 because papers can contain multiple synonyms.Source: §III · p. 3

A. AI-based systems impact on safety (RQ.1)

The review coded most studies as viewing AI as reducing AV safety risks, while a smaller group treated AI as a potential threat. These groups focused on different topics, with risk-oriented work emphasizing safety assurance.

A. AI-based systems impact on safety (RQ.1)

  • 81% (48) of the papers were coded as arguing that AI decreases safety risks, while 19% (11) considered AI a potential threat to system safety.The review’s topic grouping follows these two impact categories.Source: A. AI-based systems impact on safety… · p. 3

B. Main topics of the studies (RQ.2)

  • The decrease-risk group covered Sensors and Perception (21 papers, 44%), Navigation and Control (13, 27%), Fault Prevention (6, 13%), Conceptual Model and Framework (4, 8%), and Human Factor (4, 8%).The potential-threat group covered Fault Forecasting (5 papers, 45%), Ethics and Policies (4, 36%), and Dependability and Trust (2, 18%).Source: B. Main topics of the studies (RQ.2) · p. 4
  • Sensors and Perception studies addressed environmental detection and navigation support, while Navigation and Control studies addressed autonomous navigation and vehicle control.Sensors and Perception was the largest topic with 21 studies; Navigation and Control had 13.Source: B. Main topics of the studies (RQ.2) · p. 4
  • Fault Prevention studies addressed fault occurrence, security, diagnosis, maintenance, and prognosis, while Human Factor studies covered driver behavior, transitions, monitoring, and interfaces.Four studies each addressed Human Factor and Conceptual Model and Framework; the latter included safety, reliability, behavioral uncertainty, and highway driving.Source: B. Main topics of the studies (RQ.2) · p. 4
  • Fault Forecasting studies addressed evaluation, testing, verification, and validation, while Ethics and Policies and Dependability and Trust focused on ethical decisions, certification, safety, and robust operation.Ethics and Policies papers included one study of ethical frameworks in crash decisions; Dependability and Trust studies addressed safety issues and robust operation with non-deterministic software.Source: B. Main topics of the studies (RQ.2) · p. 5

C. Techniques used (RQ.3) and problems they seek to address (RQ.4)

The review classified reported AI techniques and the problems they addressed, then synthesized the studies’ findings across research topics. Reported results ranged from positive experimental applications to theoretical proposals and preliminary work.

C. Techniques used (RQ.3) and problems they seek to address (RQ.4)

  • The review mapped sample papers into technique and problem categories based on how their authors described the AI used, adapting category terminology from field literature [13][14][15][16][17].Some papers reported only general AI or ML approaches rather than specific techniques: 14 papers (24%).Source: C. Techniques used (RQ.3) and problems… · p. 5
  • 22% (13) of the studies used artificial neural network techniques, and 17% (10) used SVM.Other reported techniques included Fuzzy Logic (8%, 5), Bayesian AI (7%, 4), Hidden Markov models (7%, 4), and Estimation Filters (7%, 4).Source: C. Techniques used (RQ.3) and problems… · p. 5
  • Canny Edge Detection, case-based reasoning, k-means, complex decision trees, and several other techniques appeared only once, leaving room for studies of underrepresented approaches.The review lists additional one-time techniques including Conditional Random Fields, Gaussian Mixture Models, path-planning algorithms, and the Viterbi Algorithm.Source: C. Techniques used (RQ.3) and problems… · pp. 5–6

D. Reported findings (RQ.5)

  • Sensors and Perception papers reported positive results for recognition and detection tasks, including ANN-based recognition of road signals and pedestrians and SVM-based detection of roads, traffic lights, and pedestrians.The review attributes progress to developments in AI and sensor technologies.Source: D. Reported findings (RQ.5) · p. 6
  • Navigation and Control papers reported positive results, while Fault Prevention papers reported promising results for several cyber-attack detection approaches; some Fault Prevention work remained preliminary or untested.A hybrid architecture was tested for precrash control, and SVM supported safer path planning around obstacles. Fault Prevention examples included ANNs, Estimation Filters, and Fuzzy Logic.Source: D. Reported findings (RQ.5) · p. 6
  • Human Factor findings were preliminary or theoretical, while conceptual frameworks generally lacked experimental results and papers on assurance, ethics, and trust mostly proposed future solutions.Three exceptions included simulated crash scenarios under ethical theories, a preliminary test-case approach, and a Bayesian Deep Learning architecture without real-experiment findings.Source: D. Reported findings (RQ.5) · p. 6

E. Reported future studies (RQ.6)

The reviewed studies proposed future work to address research limitations and extend their approaches. Suggested directions included broader testing, real-world evaluation, improved vehicle dynamics and sensing, and safety analysis of system behavior.

E. Reported future studies (RQ.6)

  • Sensors and Perception studies proposed testing under low contrast, severe weather, and nighttime conditions, as well as reducing computational costs and using virtual environments for testing.Examples included brake-lamp recognition, traffic-light detection, speed-bump detection, and parallel image-recognition computing.Source: E. Reported future studies (RQ.6) · p. 7
  • Navigation and Control proposals included modeling full vehicle dynamics, predicting obstacle margins ahead of time, improving sensor fusion when GPS is lost, and simulating more vehicle and intersection types.The margin extension was proposed for the SVM approach to avoid maneuvers too close to obstacles.Source: E. Reported future studies (RQ.6) · p. 7
  • Fault Prevention studies proposed real-vehicle cyber-attack testing, broader security evaluations, additional prognosis methods, and analysis of the safety implications of CPU-use underestimation.The CPU-utilization work also proposed automatic feature selection and investigation of how to handle underestimation.Source: E. Reported future studies (RQ.6) · p. 7
  • Ethics and Policies papers called for more research, while a Human Factor study proposed investigating rear-end collision mechanisms and a framework study suggested large-scale resilience and sustainability research.Half of the Conceptual Model and Framework studies did not suggest future work; experimental deployment was described as an implicit next step.Source: E. Reported future studies (RQ.6) · p. 7
  • Fault Forecasting papers proposed safety-envelope research, end-to-end design and validation, automated test cases, and broader traffic-safety simulations with statistical analysis [67][66][68].Dependability and Trust studies called for concrete safety metrics and investigation of interdependencies among people, systems, and environments.Source: E. Reported future studies (RQ.6) · pp. 7–8

IV. SLR FINDINGS ORIENTED BY AN AV SYSTEM MODEL

The review maps AI-and-AV-safety studies onto an adapted semi-autonomous system model to show where research is concentrated and where safety-related gaps remain. Findings cover the model components, AI techniques, and problems addressed by the selected studies.

Adapted DDT model

  • The adapted model groups human-related elements into a human-in-the-loop component interacting with machine actuators, control, perception, and the environment.The authors use it to map literature coverage and identify areas with no studies.Source: §IV · p. 8

Model-mapped findings

  • Only system-wide studies considered both increased and decreased safety; component-focused studies identified increased safety risk, and none addressed the vehicle, machine actuators, or environment.Studies were concentrated in machine perception, followed by broad system views, machine control, and human-in-the-loop aspects.Source: §IV · p. 8

Technique mapping

  • AI techniques were grouped as system-oriented (32%, 19) or component-oriented (68%, 40), with technique patterns differing across perception, control, and human-in-the-loop studies.General AI/MI was common in system-wide studies; ANNs in perception; fuzzy logic in control; and Bayesian AI in human-in-the-loop studies.Source: §IV · p. 9

Technique heterogeneity

  • 24% of papers used generic AI/ML concepts, 49% used one technique, and 27% combined multiple techniques.Among single-technique studies, the most frequent were ANN (28%, 8), fuzzy logic (14%, 4), and SVM (10%, 3).Source: §IV · p. 9

Hybrid approaches

  • Seven papers combined ANNs with other techniques, and the review also found combinations involving SVM, hidden Markov models, Hough transformation, regression, and image recognition.The reported combinations span navigation and control as well as sensors and perception.Source: §IV · pp. 9, 12

Problems addressed

  • Among component-level problems, 21 papers (36%) concerned machine perception, 17 (22%) machine control, and 7 (10%) the human-in-the-loop DDT component.The review also identifies 16 system-level problems, including AV validation and human-machine coexistence.Source: §IV · pp. 10–11

V. FINAL REMARKS

The review finds an immature evidence base for AI and AV safety, with practical results concentrated in machine perception and many findings still preliminary. It concludes that future research needs a stronger safety agenda, including validation, human-machine collaboration, and assurance.

Evidence and maturity

  • Only machine perception had more studies with practical results; few studies in other components reported real deployments, and the authors characterize the field as not mature.Only 24% of the papers considered were published in journals.Source: §V · p. 12
  • Navigation and control studies reported promising results, but their achievements were less mature than perception results and most were preliminary.Source: §V · p. 12

Human factors and scope

  • Human-in-the-loop research needs to examine people as sources of safety-relevant interactions, including how humans and machines reach or fail to reach shared decisions.The review also calls for immersive environments for AV development and testing, deeper ethical analysis, and more autonomous-truck studies.Source: §V · pp. 12–14

Safety gaps

  • The review found few studies treating fault forecasting, ethics and policies, and dependability and trust as potential negative impacts of AI on safety.The authors connect this gap to a perceived weak influence of safety-engineering culture in the field.Source: §V · p. 12

Research needs

  • Most reviewed studies need testing in simulated or real settings, broader scenarios and datasets, and comparisons with benchmarks or alternative techniques.The authors also question whether reported error rates suit real AV deployment without hazard analysis.Source: §V · p. 12

Safety agenda

  • The proposed research agenda includes machine-learning validation, autonomy assurance and trust, ethical decisions, and deeper study of human-machine collaboration.The authors identify validation and assurance as important to SAE level 5 and safety certification of non-deterministic control systems.Source: §V · pp. 12–13

APPENDIX

The appendix summarizes diverse studies and proposed approaches concerning autonomous-vehicle safety, spanning system design, validation, control, monitoring, verification, perception, prediction, sensor fusion, and security. It presents reported findings alongside limitations, operating conditions, and unresolved challenges.

APPENDIX

  • Three proposed architectures address safety failures: two target human ignorance and AV behavior, while a third secures vehicles after system failure; all depend heavily on connectivity.The architectures are described as resource-dependent.Source: APPENDIX · p. 18
  • Soft classifications should propagate to the decision layer so AVs can act more cautiously; interpretability, compliance, and clear component metrics support trust and performance evaluation.Safety, interpretability, and compliance are identified as themes for AV adoption; the discussion argues that uncertainty should influence decisions rather than be hidden by hard predictions.Source: APPENDIX · p. 19
  • Validating machine-learning systems to the ultra-dependable levels required for AV fleets remains a challenge; a proposed certification strategy spans engineering, software, security, testing, social, and legal concerns.The paper discussing this validation challenge offers no particular finding; the certification strategy also covers hardware reliability, robotics, and human-computer interaction.Source: APPENDIX · p. 19
  • Because AV virtual assistants lack common frameworks for ethics, liability, data protection, privacy, and security, the discussion recommends initially limiting their autonomy to safety-related functions.It cautions that fully autonomous assistants could create more drawbacks than benefits, while noting that user experience remains uncertain.Source: APPENDIX · p. 19
  • A proposed data-analysis framework combines AI, quantified self, IoT, and automated driving to model driver behavior and reduce prediction uncertainty, which remains difficult to forecast amid fluid interacting factors.The framework is intended to monitor, predict, and control a transportation system.Source: APPENDIX · p. 20
  • An autonomous-highway-vehicle concept is proposed to use vehicle autonomy to benefit drivers and improve highway driving safety.Source: APPENDIX · p. 20
  • AVEthics is proposed as an initial framework for AV AI ethics policy, addressing decisions that may harm agents and should respect end-user and stakeholder interests.The discussion says such a policy is needed and characterizes the proposed framework as the beginning of a longer process.Source: APPENDIX · p. 20
  • In specific crash scenarios, utilitarian ethics may reduce fatalities, while other ethical systems may inform machine-learning rules; the experiment remains a work in process.The experiment integrates utilitarianism, respect for persons, and virtue ethics in a crash scenario involving three crash types.Source: APPENDIX · p. 20
  • Increasing highly automated vehicle penetration with suitable parameter adjustment may improve traffic safety, and microscopic simulation can compare safety and efficiency across scenarios.Source: APPENDIX · p. 21
  • A Bayesian transition model illustrates real-time switching from assisted to automated driving when collision probability is high, while derived LTL test cases can exercise continuous vehicle behaviors.Traffic recordings can support automatic test-case derivation and machine-learning safety certification, and the approach can be extended to additional traffic situations.Source: APPENDIX · p. 21
  • Driver-monitoring research identifies non-contact physiological assessment as challenging, while preliminary findings support skeletal-data maneuver classification and report visible-light classifiers outperforming thermal infrared classifiers, which performed acceptably across weather conditions.SVM was also reported as the best-performing classifier for prognosis of ADAS safety-critical component remaining useful life.Source: APPENDIX · pp. 22, 25, 26
  • A low-cost remote-controlled car prototype showed efficient road detection on clearly marked lanes, while Canny edge detection had a low error rate.A separate NEAT driving experiment learned soft-curve turning within 50 generations but still struggled with sharp curves and sometimes crashed into track limits.Source: APPENDIX · pp. 22–23
  • A public demonstration completed a 100-km urban and motorway route autonomously, with CACC distance error below 1 metre, but a 900-metre tunnel required deactivation.The system also handled motorway overpasses and traffic signals; the authors noted that localization needed improvement for longer GPS gaps.Source: APPENDIX · p. 23
  • A 65 percent reduction in total intersection delay was reported for four crossing vehicles using the proposed algorithm, compared with AWSC intersection control.The study was described as being in its initial stages.Source: APPENDIX · p. 24
  • Path-tracking experiments reported improved comfort and safety through geometry-informed speed control, while a Big Data method proposed safest-optimal trajectory selection using accident, ETA, and clearance-time predictions.The MMS-CH algorithm was reported to handle changing dynamic-environment data while avoiding unnecessary recalculation.Source: APPENDIX · pp. 24–25
  • Formalization and Z3 verification demonstrated safety for a vehicular coordination case study; responsibility-based coordination was judged suitable for safety reasoning, though verification may become harder as model detail and scale increase.Source: APPENDIX · pp. 24–25
  • GPU pedestrian detection ran six times faster than CPU detection, while traffic-sign and traffic-light recognition ran two to three times faster than pedestrian detection.Detection was difficult in dark conditions or backlight; separate work reports that 3D LIDAR pedestrian detection can operate in varied lighting, and virtual-world training with adaptation improved detectors while saving about 90% of annotation effort.Source: APPENDIX · pp. 26, 32
  • AdaBoost generalized without overfitting for road-junction classification, while CRF and HMM post-processing reduced classification noise and clarified junction boundaries.Other perception studies reported human-centric ranking of surrounding-object importance and medium-level fusion that preserved low-level information while enabling real-time processing.Source: APPENDIX · pp. 26, 27, 29
  • LSTM and GRU used sequential information to detect and identify sensor attacks, with LSTM achieving the highest accuracy among the data-based methods and both models excelling on coincident attacks.GRU was faster to calculate than LSTM, but the passage says CPU use on a general computer was not a barrier to detecting sensor attacks.Source: APPENDIX · p. 26
  • A turn-signal classifier achieved 95% accuracy, and its sequential-driving frequency recognition achieved 82.2% accuracy; another system used tracking and forecasting to stabilize traffic-light detection.A road-sign approach classified shapes with an MLP and sign contents with an SLP, reporting better performance than most compared approaches and relatively real-time software execution.Source: APPENDIX · pp. 26, 27, 31
  • An LSTM pedestrian predictor outperformed a constant-velocity Kalman filter even with small datasets and adapted to changed walking paths using only a small portion of new data.For trajectory classification, GMM performed best in low noise but was judged impractical for real environments; CHMM suited turning trajectories and DHMM straight trajectories.Source: APPENDIX · pp. 28, 30
  • Particle filtering detected 89% of spoofed beacons, compared with 76% for Kalman filtering, while leaving missed rates of 11% and 24%, respectively.Particle filtering handled both linear and nonlinear paths; Kalman filtering performed well on linear roads but lagged on nonlinear paths. The passage also reports computational-overhead savings of 85.94% for particle filtering and up to 86.5% for Kalman filtering in the same scenario.Source: APPENDIX · p. 28
  • Median 5-minute-bin vehicle-count error was 0.0686 for the proposed method, versus 0.0957 for multiple-target motion tracking and 0.2290 for VDL.The evaluation used a dataset with over 88 hours of video; the proposed method also had fewer and less severe errors across the testing set.Source: APPENDIX · p. 30
  • A 3D-LIDAR road-area classifier using SVM and HOG features reported an 8.51% error rate and correctly identified road up to 10 m ahead; a lane-detection prototype was validated in preliminary scene tests.A separate Doppler-profile collision and driving-scenario classifier was proposed as a potential V2V collision-avoidance service, but no experimental results were presented.Source: APPENDIX · pp. 30–31
Loading 1904.02697v1…