A multi-objective dynamic detection model in autonomous driving based on an improved YOLOv8

Abstract

How to efficiently identify and accurately track multiple targets in complex traffic scenes has become a key technical challenge that urgently needs to be solved. Therefore, this study developed a lightweight model that combines attention mechanism with convolutional neural network and YOLOv8 to achieve dynamic detection algorithm for multiple targets. Firstly, in the data preprocessing stage, mosaic image enhancement technology is used to highlight the features of small targets and complex scenes. Subsequently, MobileNetV3_CA network was designed to integrate lightweight MobileNetV3 and Coordinate Attention module as the backbone feature extraction network of YOLOv8, thereby enhancing the accuracy and specificity of feature extraction. To further enrich feature information, multi-scale feature layers are input into PANet for fusion. The experimental results show that the proposed method, which combines attention convolution and YOLOv8 for multi-target dynamic detection in autonomous driving scenes, significantly improves performance in multi-target detection tasks. Compared to the traditional YOLOv8 model and its variants without attention mechanism, this method outperforms the detection accuracy on multiple autonomous driving datasets. This research achievement provides a strong technical support for the perception module of the auto drive system, and helps to promote the development of autopilot technology to a higher level.