Research On Unpiloted Image Segmentation Algorithm Based On Deep Learning

In recent years,with the development of technologies and the opening policies,unpiloted driving has become one of the most rapidly developing artificial intelligence industries.Vision-based perception is the basic part in the driving system.Fisheye images are widely used because of the large viewing angle and rich information.Based on the technologies of deep learning,a semantic segmentation(pixel level)algorithm and a instance segmentation(instance level)algorithm are designed and implemented with fisheye images.The main work contents are as follows:1.Construct a fisheye image segmentation dataset for unpiloted driving.The data are collected from parking lots and city roads in Beijing and Shanghai.It take about 180 hours of fisheye camera video(frame rate:30FPS,resolution:1280*720),in which 12,000 fisheye images are extracted and manually labeled with LabelMe to form 6,000 images of semantic segmentation and instance segmentation respectively.2.For the pixelwise unpiloted scene understanding requirements,a semantic segmentation algorithm based on layer cascading is proposed.The network is designed with two different depth convolutional layer branches,in which the pixels that correct classification confidence above the threshold are input into the shallow branch,and the pixels below the threshold are input into the deep branches to implement the learning mode of layer cascade.In addition,the deep network branch utilizes the method of dense feature aggregation to get the segmentation result.The experiments show that compared with the mainstream method FCN(Fully Convolutional Network),the mIoU increased by 10.4%to 76.5%,and the test speed increased by 2.1 times to 20.2 FPS.3.For the problem of semantic segmentation can not distinguish different instances,a instance segmentation algorithm based on improved SSD(Single Shot MultiBox Detector)is proposed.The network is divided into three modules:object detection,post-detection processing,and mask segmentation.In the object detection module:designing a lightweight basic network and adding the circular scrolling convolution in the multi-scale prediction network.The second module corrects the detection result by using multi ROI Align.In the last module,the mask labels are generated dynamically and the instance segmentation is implemented with the shallow downsampling-upsampling network.Experiments show that compared with the mainstream algorithm Mask R-CNN,AP is increased by 4.34%to 41.15%,and the test speed is increased by 1.92 times to 9.8 FPS.