Autonomous vehicles use various sensors such as radar, LiDAR and GPS, along with computer vision algorithms, to understand their environment.These sensors gather data that needs to be analyzed for obstacle detection and navigation. However, achieving accurate object recognition is difficult due to challenges in data processing, high computational needs, and memory requirements .This study proposes a modified structure of MobileNet , called MobileNet-Att, which includes two attention mechanisms: Parallel Convolution Block Attention Module (PCBAM) and Squeeze-and-Excitation (SE) blocks. PCBAM captures multi-scale spatial features using parallel convolutions, enabling the model to focus on varying levels of spatial information. This design improves object classification and efficiency without increasing computational costs by effectively capturing richer contextual information. In the next step, SE blocks readjust the importance of each channel by "squeezing" global information through average pooling, and then "exciting" the channels based on this global context. This enables the network to emphasize essential features while minimizing the influence of irrelevant data. In essence, MobileNet-Att, with its attention mechanisms and modifications, offers a balanced approach between performance and computational loading to provide a valuable solution for object classification in autonomous vehicles. Experiments show that MobileNet-Att outperforms earlier models in accuracy and parameter efficiency on the CIFAR-10 and Caltech-101 datasets.