Road Targets Detection Based On Convolutional Neural Network Algorithms

As core branch filed of computer vision,The main task of target detection is to find the targets of interest in a given input image and determine their accurate size and positions.Target detection can often be interfered by shape,posture,size,appearance,lighting environment,and occlusion of other target objects under real circumstance.Feature extraction of the traditional detection algorithms often require complicated process,while deep learning can learn millions of parameters from input images automatically,and makes a huge leap performance in image classification,recognition,detection and other related fields.This dissertation mainly studies the road targets algorithms based on convolutional neural network.The main content of the dissertation are listed as follows:(1)Method based on feature fusion for Faster-RCNN targets detection is proposed.Faster-RCNN,as a classic target detection algorithm,has superior detection performance in ImageNet computer vision competition.Based on the Faster-RCNN detection algorithm,the pyramid feature fusion processing is applied to the base network of the proposed feature,and then deconvolution operation is applied to achieve the fusion of the global and local features of the target region,so that the performance of the model features can be improved.The KITTI dataset used in our experiment was divided into three categories: simple,median-difficult and difficult according to the targets occlusion region.Finally,the experiment showed that our algorithm achieved from 1% to 4% accuracy for road motor vehicles and pedestrians.(2)The deep learning task generally chooses Batch Normalization as the data normalization method,and Batch Size generally takes a value of 32 or 64.However,due to the large amount of computation,the original Batch Size would excess the the general computer computing power,while the smaller Batch Size would degrades the performance.The group normalization used in this paper can ensure that the lower Batch Size will not have big impact on the detection performance.Data augmentation adopted image blending and is tested on KITTI public datasets.The result shows that the image blending augmentation and Group Normalization used in this paper also have a certain improvement on the pedestrian detection of the SSD target detection algorithm.The improved algorithm proposed in this paper has been verified by a series of experiments on the KITTI dataset,which showed the proposed algorithm can improve the targets detection performance.