The development of autonomous cars faces challenges in sensor dependability despite the substantial advances in sensing technology. Convolutional Neural Networks (CCNs) is an important algorithm in identifying object detection obstacles in decision-making. In addition, they excel in image processing and object detection tasks due to their trainable parameters and gradient descent approaches. The goal of this thesis project is to enhance object identification in autonomous cars in unexpected conditions, namely identifying construction zones on roadways to improve safety and reliability. Using a pre-trained MobileNetV2 SSD model, the research examines several learning rate schedulers and optimization approaches implemented in Pytorch, including StepLR, MultiStepLR, ExponentialLR, CosineAnnealingLR, SGD, and ASGD. The study reveals that while simpler approaches like StepLR and MultiStepLR outperform others in terms of training loss reduction, prioritizing simplicity and ease of implementation, more complex techniques like CosineAnnealingLR and ReduceLROnPlateauLR offer advantages in specific scenarios. Moreover, ASGD consistently outperforms SGD in convergence speed and validation accuracy, underscoring the importance of selecting appropriate optimization methods tailored to the task at hand.