Enhancing Autonomous Vehicle Navigation Using Deep Learning-Based Scene Understanding

To navigate complex environments, AV navigation relies on the accuracy of scene understanding. In this work, a new approach based on deep learning is brought into relation with the object recognition, scene segmentation, motion prediction, and reinforcement learning for enhancing AV navigation. The methodology starts by utilizing the scene segmentation accomplished by breaking down the drivable areas along with obstacles using the U-Net architecture. Real-time object recognition using YOLO is incorporated at the end. With an LSTM network, the AV can predict potential future behaviors of moving objects. RRT<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">*</sup> is best for collision avoidance while navigating to build up a route. Conclusion: A Deep Q-Learning reinforcement learning system improves real-time navigation and provides a framework for optimization for dynamic decision-making and vehicle control situations. Simulation and experimental results also demonstrate that the technology increases the navigation efficiency and reduces the frequency of collisions, which improves the process of obstacle avoidance. Thereby, 92.5 % precision for U-Net can be proven.