针对目前自动驾驶领域的目标检测算法在交通场景下的漏检目标，目标定位不精确、目标特征表达不充分及目标识别效果欠佳等问题，提出一种基于TPH-YOLOv5（transformer prediction heads-YOLOv5）的道路目标检测方法。首先为了减轻物体尺度急剧变化带来的漏检风险，增加了用于微小物体检测的检测头，为在高密度场景中精确定位对象，使用Transformer预测头来捕获全局信息；其次为了增强模型的特征表达能力，用选择性注意力机制（selective interactive module with affinity learn-ing, SIMAM）模块对卷积层的输出进行加权；最后，为了提高目标识别的精度，网络颈部增加了4个金字塔池化模块（spatial pyramid pooling, SPP）块来进行多尺度融合，为了加快收敛速度和提高回归精度采用EIOU（Euclidean distance-IOU）作为边界框损失函数。通过消融、对比和可视化验证实验表明，提出的算法比YOLOv5在平均精度上提高了8.1%，漏检率明显减少，目标检测效果明显增强。

Aiming at the problems of missed targets, inaccurate target localization, insufficient target feature expression, and unsatisfactory target recognition effects in the current object detection algorithms for autonomous driving in traffic scenarios, a road object detection method based on transformer prediction heads-YOLOv5 (TPH-YOLOv5) was proposed. Firstly, to reduce the risk of missed objects caused by drastic changes in object scales, a detection head for detecting small objects was added, and a Transformer prediction head was used to capture global information for precise object localization in high-density scenes. Secondly, to enhance the feature expression ability of the model, the output of the convolutional layer was weighted using the selective interactive module with affinity learning (SIMAM) module. Finally, in order to improve the accuracy of target recognition, four spatial pyramid pooling (SPP) blocks were added to the neck of the network for multi-scale fusion, and the Euclidean distance-IOU (EIOU) was used as the bounding box loss function to accelerate convergence speed and improve regression accuracy. Through ablation, comparison, and visualization experiments, it is shown that the proposed algorithm improves the average precision by 8.1% compared to YOLOv5, significantly reduces the missed detection rate, and enhances the object detection performance.