Research On Instance Segmentation For Scene Of Unmanned Driving With Deep Learning

With the development of computer technology,unmanned driving has become one of the hot topics of current research.With the rapid development of deep learning and computer vision algorithms,the intelligence of unmanned driving has been greatly improved,making it possible for unmanned driving to analyze complex scene information using low-cost images.Computer vision algorithms commonly used in unmanned driving include object detection algorithms and image semantic segmentation algorithms.Although the object detection algorithm can predict the approximate area where each instance is located in the environment,it c annot classify the pixels at the pixel level.Image semantic segmentation can provide global pixel-level classification information,but cannot distinguish between different instances.The instance segmentation algorithm integrates the prediction results of the two tasks to provide richer specific instance-related information in the environment for the unmanned scene,so that the unmanned vision technology has stronger environmental information analysis capabilities.At present,most mature instance segmentation algorithms are two-stage,with slower inference speed,and are not suitable for unmanned scenarios with high real-time requirements.This dissertation proposes a fast instance segmentation algorithm model based on a one-stage object detection algorithm.This method replaces the way of Mask RCNN object detection to the YOLOv3 algorithm directly.Generate the final bounding box,use repooling technology to extract the bounding box from the corresponding feature map and crop it to a fixed size,then pred ict the mask of the instance from these fixed size feature maps,and finally generate instance mask in the original image by upsampling operation.This model solves the problem of slow target detection in the classic instance segmentation algorithm and obtains a faster instance segmentation algorithm.Aiming at the problem that the proposed fast instance segmentation algorithm is not high-accurate and not fast enough,an attention mechanism is introduced in the image semantic segmentation feature fusion part of the mask prediction to improve the mask prediction capability.The object detection module uses a fast NMS algorithm instead of the traditional NMS algorithm accelerates the reasoning speed of the model.Through two improvements,a fast instance segmentation algorithm that can finally achieve both accuracy and speed is obtained.This dissertation uses COCO and City Scapes data sets for research related experiments.Cross-contrast experiments was carried out with set variables to determine the optimal combination of each module of the model,and compared with the current mainstream instance segmentation algorithm,the fea sibility of this method was demonstrated.Finally,the actual measurement in the campus scene proves the effect is good.