Uncertainty quantification is crucial in developing and deploying autonomous driving systems. Autonomous vehicles operate in complex and dynamic environments where uncertainties in road conditions, traffic scenarios, and sensor limitations can significantly impact their performance and safety. Accurately quantifying and managing uncertainty ensures reliable decision-making and mitigating risks associated with autonomous driving. For autonomous vehicles, it is essential to develop efficient solutions regarding memory and computation requirements. \Glspl{av} operate in real-time, dynamic environments where quick and accurate decision-making is critical. Therefore, \Gls{uq} methods must balance providing reliable uncertainty estimates and being computationally lightweight. We propose a new method to address some shortcomings of previous approaches, reducing the computational overhead of doing multiple samplings and the memory usage of having various models by generating numerous predictions using multi-model sampling techniques. Test-time augmentations can be applied to the input images at inference time to produce diverse outputs that are still plausible according to the learned distribution. They can be used instead of maintaining a large number of models or having to run multiple dropout samplings. The work in this thesis investigates the usage of augmentations, such as noise, blurring, or pixel value manipulations, on the input image as a batch prediction and using the output to estimate the uncertainty of the model. The proposed approach is tested on Yolov5 trained on the Mapillary dataset. Furthermore, we demonstrate the efficacy of the proposed algorithm in an active learning experiment. We further compare our method with two state-of-the-art techniques: deep ensembles and MC-Dropout.