MCPT: Mixture of CNN and Point Transformer for Multimodal 3D Object Detection

3D object detection is an indispensable techniques for enhancing the safety of autonomous driving systems, serving as the key to perceiving the surrounding environment in real time with high precision. In this paper, we aim to overcome the trade-off between accuracy and efficiency in conventional methods by proposing a novel module that fuses the strengths of convolution—adept at local feature extraction—with those of self-attention mechanisms that capture global interrelationships. Specifically, by incorporating a parallel 7×7 Depthwise Separable Convolution into Point Transformer, our approach effectively extracts both fine details and broad contextual information, and by replacing the image encoder with Depthwise Separable Convolution, it reduces computational load and model parameters. Experimental results confirm that our proposed method achieves a 31% reduction in the number of parameters while improving detection accuracy by 0.44%, demonstrating a balance between computational efficiency and precision. As a result, this approach is expected to contribute to the development of practical autonomous driving systems in resource-constrained edge device environments, ultimately enhancing overall safety and reliability.