TOM SEMINAR_1_ FINAL.pptx tom pdf file he used to prsent

Tom presentation file in mbccet

1. 17-10-2025 1 Department of Mechanical Engineering Technical seminar on “ Modern Vehicle Control Systems: Integrating Vision-Based Automatic Braking for Enhanced Safety ” Presented by Mr. TOM JOSE MBC22ME020 Under the guidance of Mr. AJU JOHN Assistant Professor Department of Mechanical Engineering

2. 17-10-2025 2 Introduction Problem Statement Objective Literature Survey Relevance of key Literatures to Our Project Methodology and System Architecture Results and Discussion Applications Future Research Conclusion References CONTENTS

3. I NTRODUCTION The AI Revolution in Automotive Systems: The automotive industry is undergoing a paradigm shift, leveraging Artificial Intelligence (AI) to enhance safety, efficiency, and performance. This spans from perception systems (like ADAS) to core vehicle dynamics control (like transmission systems). Two Pillars of Intelligent Vehicles: External Perception:  Using sensors and cameras to understand the environment and make critical safety decisions (e.g., Automatic Braking). Internal Control:  Using AI to optimize the vehicle's own mechanical systems for superior performance and comfort (e.g., Gear Shifting). 3 17-10-2025

4. PROBLEM STATEMENT   Traditional Automated Braking Systems rely on expensive sensor suites (LiDAR, Radar), increasing vehicle costs and limiting widespread adoption.  Control design for complex systems like automatic transmissions is a time-consuming, manual process performed by highly trained experts, leading to long development cycles. Key Question:  Can AI provide cost-effective, high-performance alternatives to traditional automotive engineering solutions? 4 17-10-2025

5. OBJECTIVE OF THE STUDY To develop a purely  vision-based Automated Braking System  using deep learning, eliminating the need for costly LiDAR . . To automate the synthesis of a high-performance  gearshift controller  using Deep Reinforcement Learning (DRL), outperforming conventional control methods. 5 17-10-2025

6. 6 LITERATURE SURVEY JOURNAL HEADING AUTHOR REMARKS 1 2 Deep Learning based Automated Braking Decision-making forAdvanced Driver Assistance System. Deep reinforcement learning for gearshift controllers in automatic transmissions . C.P Razeena, et.al Year:2025 Gerd gasiselmann , et.al Year:2022 This paper presents a vision-based automated braking system using YOLOv8 and ResNet18. The model achieves 94.4% accuracy, offering a cost-effective alternative to expensive sensor-based solutions for Advanced Driver Assistance Systems. . This work applies Deep Reinforcement Learning to control gearshifts in an automatic transmission. Agents trained in simulation were successfully transferred to a real test bench, outperforming conventional controller performance . 17-10-2025

7. RELEVANCE OF KEY LITERATURES TO OUR PROJECT "Deep reinforcement learning for gearshift controllers"  proves AI can successfully control complex real-world car parts, moving from simulation to a physical test bench. "Deep Learning based Automated Braking Decision-making"  demonstrates a highly accurate automatic braking system using only cameras, avoiding costly sensors like LiDAR. Together, these papers provide a complete blueprint for using modern AI for both perceiving the driving environment and controlling the vehicle's actions in real automotive systems. 7 17-10-2025

8. METHODOLOGY 8 17-10-2025 Study 1: Vision-Based Automated Braking System Data:  Berkeley DeepDrive (BDD100K) dataset – diverse driving images. Object Detection:  YOLOv8 identifies vehicles, pedestrians, etc. Distance Estimation:  Euclidean distance calculated from bounding boxes. Decision Making:  A pre-trained ResNet18 model is fine-tuned to make the final "Brake" or "No Brake" decision.

9. 9 STUDY 2: DRL-BASED GEARSHIFT CONTROLLER Training Environment:  High-fidelity simulation of a Mercedes-Benz 9G-TRONIC transmission. AI Agent:  Uses state-of-the-art algorithms (PPO and SAC) to learn optimal control policies. Reward Function:  Carefully designed to balance competing goals: fast shift time vs. low jerk (driver comfort). Sim-to-Real Transfer: Domain Randomization (DR):  Trains the AI on varied simulated conditions to build robustness. Domain Adaption (DA):  Fine-tunes the simulation or the AI model using real test-bench data. 17-10-2025

10. EXPERIMENTAL PROCEDURE 10 The dataset contains a diverse collection of real-world driving scenes captured from a vehicle's perspective. Images include various conditions such as different weather (sunny, overcast, rainy) and times of day (day and night). This diversity is crucial for training and testing a robust model that can perform reliably in any real-world driving scenario. 17-10-2025

11. The system processes driving images through YOLOv8 for object detection and distance calculation, then uses this annotated data to train a ResNet18 model that decides whether to brake based on object proximity. 11

12. This diagram shows how AI gear-shift controllers are first trained safely in a simulation, then two methods—"test-driven" fine-tuning and "data-driven" parameter optimization—are used to bridge the "sim-to-real gap" and successfully transfer the AI controller to a physical transmission test bench. 12

13. RESULTS The model learned quickly and effectively , with both training and validation accuracy rising steadily and converging around the 10th epoch. It generalizes well to new data , as the validation loss closely follows the training loss, showing it didn't just memorize the training examples. Training was stopped optimally  at the 10th epoch, as performance stabilized, preventing unnecessary computation and potential overfitting. 13

14. Study 1: Deep Learning based Automated Braking Decision-making Created a camera-only braking system  using AI (YOLOv8 & ResNet18), avoiding expensive sensors like LiDAR. Achieved 94.4% high accuracy  in deciding when to brake, proving vision is a reliable alternative. Offers a cost-effective safety solution , making advanced driver assistance more accessible for the future. 14 17-10-2025

15. Study 2: Deep Reinforcement Learning for Gearshift Controllers in Automatic Transmissions Successfully replaced manual tuning  with an AI controller for smoother, faster automatic gear shifts. Trained AI in simulation , then successfully transferred it to a real transmission using special adaptation methods. Outperformed standard controllers  by reducing shift time while maintaining comfort, proving AI's real-world potential. 15

16. APPLICATIONS Vision-Based Braking System: Standard ADAS:  Automatic Emergency Braking (AEB), Forward Collision Warning. Cost-Sensitive Vehicles:  Bringing advanced safety to budget-friendly cars. Robotics & Drones:  For obstacle avoidance in other autonomous systems. cars. 16 17-10-2025

17. 17 C ONTINUE… AI-Based Transmission Control: Next-Generation Vehicles:  Smoother and faster shifting in luxury and performance cars. Automated Calibration:  Drastically reducing development time and cost for new transmission models. Adaptive Control:  Systems that can adapt to component wear or different driving styles over time .

18. FUTURE RESEARCH AND ADDITIONS Enhanced Perception:  Fusing cameras with low-cost radar for all- weather reliability. Complex Scenarios:  Extending the DRL controller to handle dynamic gear shifts involving multiple clutches and engine torque control. Lifelong Learning:  Developing AI systems that can continue to learn and adapt after deployment in real vehicles. Robustness & Safety:  Formal verification and validation of AI models to ensure fail-safe operation under all conditions. 18 17-10-2025

19. 19 Vision is Viable:  The vision-based braking system demonstrates that high-performance ADAS can be achieved without expensive sensors, aligning with a cost-effective future for vehicle automation. AI Outperforms Tradition:  Deep Reinforcement Learning has proven its capability to control complex real-world systems like automatic transmissions, outperforming traditional methods and reducing engineering effort. A Paradigm Shift:  Together, these studies highlight a paradigm shift towards intelligent, data-driven systems that enhance both vehicle safety and performance, paving the way for the fully autonomous vehicles of tomorrow. CONCLUSION

20. 20 R EFERENCES C.P. Razeena et al., "Deep Learning based Automated Braking Decision-making for Advanced Driver Assistance System,"  Procedia Computer Science , 2025. G. Gaiselmann et al., "Deep reinforcement learning for gearshift controllers in automatic transmissions,"  Array , 2022. J. Schulman et al., "Proximal Policy Optimization Algorithms,"  arXiv:1707.06347 , 2017. T. Haarnoja et al., "Soft Actor-Critic Algorithms and Applications,"  arXiv:1812.05905 , 2018. 17-10-2025

21. C ONTINUE… Haus, S., Anderson, R., Sherony , R., & Gabler, H. (2021). Potential effectiveness of bicycle-automatic emergency braking using the Washtenaw Area Transportation Study data set.  Transportation Research Record . Vaiyapuri , T., Mohanty, S.N., Sivaram, M., Pustokhina , I.V., Pustokhin , D.A., & Shankar, K. (2021). Automatic vehicle license plate recognition using optimal deep learning model.  Computers, Materials & Continua . Lampe, A., Serway , R., Siestrup , L., & Guehmann , C. (2019). Artificial intelligence in transmission control – Clutch engagement with reinforcement learning. Deuschl , K., Doster, T., Gaiselmann , G., & Studer, S. (2020). Automated functional development for automatic transmissions using deep reinforcement learning.  ATZ Electron Worldwide, 15 , 8–13. 21

22. THANK YOU 22 17-10-2025