Vehicle detection is a challenging problem in autonomous driving systems, due to its large structural and appearance variations. In recent years convolutional neural networks (CNN) have achieved great success. Presently the state-of-art R-CNN and YOLO architectures are explored, implemented, and adopted for performance comparison on real-time data. From the experimental results, the R-CNN has faster detecting speed and accuracy in complex scenes. However, the transfer learning-based CNN model can be further analyzed and validated over other datasets and can be carried out. In comparison to the state-of-art, the proposed transfer learning approach may accurately regress the vehicle shape and classify vehicle fine-grained categories. Presently, the performances are validated over the UA-DETRAC datasets. The obtained performance strongly suggests the RCNN and has better classification accuracy. The proposed transfer learning approach can be a better alternative competing to the state-of-art approaches.