Enhancing RetinaNet for Object Detection in Autonomous Driving with Limited Data

The task of object detection in such environments is especially challenging given the limited data and complex surroundings. This paper introduces an enhanced RetinaNet algorithm designed specifically for object detection in autonomous driving scenarios with limited data. The proposed algorithm incorporates two key techniques: transfer learning and the DimensionDrop FPN (DD-FPN) architecture, to overcome these challenges. Specifically, a pre-trained ResNet-50 network is fine-tuned on the COCO dataset. Following this, it is fine-tuned on a smaller dataset of autonomous driving scenarios, which helps to prevent overfitting and preserve pre-trained weights. The DD-FPN structure reduces the dimensionality of feature maps at various scales, producing FPN feature maps of different scales to achieve precise object detection. Experimental results demonstrate that the proposed algorithm produces higher accuracy and robustness than existing methods on different datasets. Therefore, this study is a valuable contribution to the development of object detection algorithms for autonomous driving scenarios with limited data and has implications for the domain of intelligent transportation systems.