For the issues of overlapping obstacle targets on driving roads, inconsistent scale characteristics, and high misdetection rates due to easy confusion with the background leading to low detection accuracy, an obstacle detection algorithm ADMH-YOLOv8 was proposed. Based on the YOLOv8 framework, firstly, a variable kernel convolution was used to improve the feature extraction network’s standard convolutional layers to enhance its feature extraction capability. Secondly, combining dilated convolution to improve the dynamic sampling operator (DySample) instead of fixed upsampling enhances the effectiveness of feature fusion. Again, an improved hybrid attention module that integrates multi-scale features was introduced to reduce information loss during upsampling and enhance the perception ability of important regions. Finally, a parameter sharing detection head (PSDH) was designed to reduce the number of parameters. The experimental results show that the ADMH-YOLO algorithm has excellent comprehensive performance, with an average accuracy of 93.8%, which is 2.0% higher than the baseline algorithm. It has high obstacle recognition accuracy and is suitable for obstacle detection tasks on roads.