LKStar-Yolov8n: an autonomous driving object detection algorithm based on large convolution kernel star structure of Yolov8n

In autonomous driving, vision-based object detection algorithms are widely used for environmental perception. However, the performance limitations of onboard devices restrict the scale of these algorithms, necessitating both high accuracy and real-time capability. To address this, we propose a new object detection network based on YOLOv8n, named LKStar-YOLOv8n. We designed a LKStar module to replace the C2f module in YOLOv8. The LKStar module utilizes reparameterized large-kernel depthwise convolutions (DWConv) and a star-shaped structure, which enhances the model's receptive field and representation capacity. Additionally, we replaced SPPF with SimSPPF to provide multi-scale feature aggregation while ensuring model speed. Evaluated on the KITTI dataset, our improved network achieved an average accuracy of 85.4%, 3.2% higher than the baseline model. We also investigated the impact of different-sized depthwise convolution kernels and the positioning of star-shaped operations on the model, demonstrating the effectiveness of our proposed architecture.