In response to the challenges posed by high-speed vehicle movement and the complex and dynamic surrounding environment, accurately identifying and locating objects of different scales such as cars, pedestrians, and traffic signal lights is not an easy task. This paper proposes an automatic driving multi-scale object detection method called GSM-YOLO. Firstly, by introducing a small object detection layer, the detection performance for small objects is enhanced. Subsequently, a GSE-SPPF module is designed to increase the feature extraction capability of the main network. Furthermore, in the neck network, the SC-ELAN structure is proposed to efficiently utilize spatial and channel information among features, capturing target information of different scales and reducing information loss. Finally, the MPDIOU function is utilized to optimize loss, improving the convergence speed and accuracy of the network. Experimental results show that on the autonomous driving BDD100K dataset, the GSM-YOLO algorithm outperforms the YOLOv5s algorithm, achieving an average precision increase of 6.1% to reach 55.7%. Additionally, the detection speed reaches 111.1 frames per second. The proposed improvement method effectively realizes object detection in road environments and can be widely applied in research related to autonomous driving object detection.