Abstract: With the innovation and development of unmanned driving technology, 3D object detection technology has entered our sight. Compared with traditional lidar-based and monocular-based 3D target detection algorithms, although binocular vision-based detection techniques cost effective, its detection effect still needs to be improved. Therefore, this paper proposes an F R-CNN 3D object detection algorithm based on an improved 3D region convolutional neural network algorithm (Stereo RCNN). The algorithm has added a frequency domain channel attention module (FcaNet) to the feature extraction network of the Stereo R-CNN algorithm, so that the model pays more attention to the semantic information related to the target from the perspective of feature diversity, and reduces the weight of the deep residual network. The impact of changes will enhance the feature extraction capabilities of the network. At the same time, unified dynamic sample weighting strategy is introduced, and the loss weights between multiple tasks are reasonably allocated during the training. While paying attention to the importance of "difficult" samples, it also considered the contribution of "simple" samples to extract more comprehensively key feature information of the object. Experimental results show that the improved F R-CNN 3D target detection algorithm has improved the average accuracy of 3D target positioning by 3%, compared with the Stereo-RCNN algorithm, and the average accuracy of 3D target detection has increased about 2%.