<p>Autonomous vehicles represent a significant advancement in the field of transportation, holding the promise of enhanced mobility and heightened safety. This study introduces a novel real-time methodology for obstacle avoidance in autonomous vehicles through the refinement of the Xception architecture, used for the first time, to our knowledge, in this field. It includes a developed lightweight deep learning model, which will enable the use of a computationally efficient model thereby reducing response time. This involves fine-tuning the model by adjusting the output layer and updating the original design’s weights. A dataset comprising 30,336 images is generated by capturing data from three cameras strategically positioned within the vehicle based on the Virtual Simulation platform for Autonomous Vehicles (VSim-AV). This training dataset encompasses diverse environmental factors and different obstacles, enhancing the model’s ability to adapt to real-world driving scenarios. In order to enrich the generated dataset, data augmentation techniques and pre-processing procedures are implemented. The developed model is first trained and then validated across different real-time scenarios, using the autonomous mode of the VSim-AV simulator. Obtained testing accuracy rate achieves 96.7%, while the average reaction time is 0.15 s.</p>