3D object detection is a critical technology for automatic driving perception systems, which accurately detects the state of the driving environment and ensures the safety of driving. Aiming at the problem of low 3D detection accuracy of small objects such as pedestrians and cyclists, a 3D object detection algorithm based on multi-scale features and adaptive non-maximum suppression (ANMS) is proposed. Firstly, a multi-scale feature extractor is designed to obtain large, medium, and small-sized features. Secondly, a multi-scale detection head is constructed to generate the candidate boxes of objects of different sizes, thereby supplementing the candidate boxes of small objects. To balance the number of multi-scale candidate boxes, a candidate box screening algorithm on ANMS is designed, enhancing the detection accuracy of objects of different sizes. The results on the KITTI dataset indicate that the improved algorithm achieves 62.57% and 73.30% detection accuracy for pedestrians and cyclists, which is 2.04% and 1.33% higher than the baseline while ensuring the detection accuracy of car-type objects, which verifies that the improved algorithm has preferable 3D detection performance in small object detection.