Compared with images, point clouds contain more accurate position information, which is widely used on environmental perception of autonomous driving. In the process of perception, due to the complexity of the scene and the sparseness of point cloud, the results of recent methods that using a single detection head are not good. In addition, the huge differences between targets, not only different categories, but also different states of the same category, also brings challenges for detection. In this paper, we designed multiple detection heads with different groups, let different detection heads detect specific category while ignoring other categories, so that different detection heads can complete different tasks. According to the target shape and the number of target categories, we set different detection heads for different samples. Noted that the categories with similar target shape are divided into the same group. Experiments demonstrate that this method can improve the weighted Mean Average Precision (mAP) of the targets and reduce the false detection. Taking into account the difference in the number of targets in different categories, the method of dividing different categories into different groups with similar number is better than that based on shape grouping.