Among the varIoUs sensors used in assisted and automatic driving systems, the perception performance of cameras and lidar is greatly affected by the weather, while the automotive millimeter-wave radar is a low-cost and almost weather-free all-weather working device.Moving objects can extract rich Doppler information. With the development of radar technology and open-source labeled datasets, object detection based on underlying radar data has become a very promising field. To address the issue of inaccurate target detection and positioning caused by low angular resolution of onboard millimeter wave radar data, and to improve the performance of millimeter wave radar target detection, this paper proposes a 3D convolutional millimeter wave radar target detection method based on prior box distance constraint to achieve the detection and classification of multiple dynamic targets. In our method, we design a feature extractor for 3D ResNet to characterize target information in Range Azimuth Doppler tensor, solving the problem of insufficient representation in existing models due to ignoring Doppler information from the original 3D radar signal; Secondly, an absolute distance Loss function is added to train the model to overcome the influence of distance on target presentation and improve the accuracy and robustness of target detection; In addition, a method of resetting prior boxes based on distance unit intervals has been proposed to solve the problem of unreasonable prior box design in existing methods.The proposed model is trained and tested on the RADDet dataset. The experimental results show that compared with the current state-of-the-art method, our model achieves the best when the IoU threshold is 0.1, 0.3, 0.5, 0.7, where IoU The improvement is the most significant when it is 0.1 and 0.3, which are increased by 6.6% and 5.1% respectively.