Given the changes of large camera angles and lighting conditions， the traditional algorithm of images’ feature point extraction cannot robustly realize extraction and matching between series images in the process of optical navigation during landing of the Mars rover. A joint detection-description feature extraction method based on convolutional neural network（CNN） is proposed to solve this. First， the video of the simulated Mars rover landing process is obtained by Blender. Then， the sparse reconstruction method is used to deal these images and real images of Mars， and establish a training dataset. Second， a convolutional neural network is built to process images by combining the dual roles of feature descriptor. And feature detector， and more accurate matching results are obtained by improving the loss function. Experiments show that this method has better results in feature detection of the Martian surface with multiple viewing angles and changing lighting conditions， and achieves better performance than traditional methods in the matching stage.