To improve the precision and robustness of 3D object detection based on stereo vision, a novel 3D object detection algorithm based on iterative self-training is proposed. To acquire the precise object point clouds for 3D object detection task, a disparity estimation algorithm based on iterative self-training is first proposed, which is capable of improving the disparity accuracy of object region by increasing the supervised signal in object region iteratively and introducing a selective optimization strategy. Then a self-adaptive feature fusion mechanism is proposed in network architecture, which adaptively fuses the features from multimodal information to obtain the precise and robust object detection results. Compared with the recent and popular algorithms based on vision system, the proposed 3D object detection algorithm achieves a great improvement in precision.