MonoICT: A Monocular 3D Object Detection Model Integrating CNN and Transformer

In the field of autonomous driving, 3D object detection is a crucial technology. Visual sensors are essential in this area and are widely used for 3D object detection tasks. Recent advancements in monocular 3D object detection have introduced depth estimation branches within the network architecture. This integration leverages predicted depth information to address the depth perception limitations inherent in monocular sensors, thereby improving detection accuracy. However, many existing methods prioritize lightweight designs at the expense of depth estimation accuracy. To enhance this accuracy, we propose the Pseudo Depth Feature Extraction (PDFE) module. This module extracts features by fusing adaptive scale information and simulating disparity, leading to more precise depth predictions. Additionally, we present a hybrid model that combines Convolutional Neural Networks (CNN) and Transformer architectures. The model employs diverse feature fusion strategies, including depth-guided fusion (DGF) and a Transformer decoder. It also utilizes a Convolutional Mixed Transformer (CMT) encoder to enhance the representation of both local and global features. Building on these innovations, we developed the MonoICT network model and evaluated its performance using the KITTI dataset. Our experimental results indicate that our approach is competitive with recent state-of-the-art methods, outperforming them in the pedestrian and cyclist categories.