The invention discloses a method for improving instance segmentation based on an unmanned driving technology. According to the method, targets are detected and classified on the basis of the Faster R-CNN based on the MASK R-CNN, and then instance segmentation is achieved through FCN feature coarse extraction and CRF optimization output. The method comprises the following specific implementation steps: step 1, classifying targets by using a partial supervision method; step 2, adopting depth separable convolution in the semantic segmentation convolution process to obtain features; and step 3, performing feature fusion optimization on the features obtained by the convolutional layer, introducing semantic information in a low layer, and introducing spatial information in a high layer. According to the method, relatively good target detection and classification results are established at relatively low cost. The method adopts depth separable convolution, thus improving the precision of a segmentation result and the efficiency of a computer, and reducing time loss.