Vehicle Detection And Segmentation Based On Convolutional Neural Network

The development of the future society will make people's lives more and more intelligent,and the intelligent transportation system will gradually integrate into people's lives.It plays an important role in facilitating people's travel and providing better transportation services.The emergence of convolutional neural networks has led to the development of computer vision,which provides the necessary conditions for the development of intelligent transportation systems.Vehicle detection and segmentation is a key step in intelligent transportation systems.The traditional detection method uses the method of scale-invariant feature transformation to extract features and then classify the method.Because the generalization ability of the model is not strong,there is a problem that the detection accuracy is not high.The full convolutional neural network R-FCN has achieved good results in the field of image detection.However,when the detected object has occlusion,deformation and small target,it is not well detected.Faced with the above problems,this thesis designs a new vehicle detection framework based on the full convolutional neural network,which can make the detection structure more accurate.The main research contents are as follows:For the case where the target size differs greatly in the same scene,it is difficult to detect.Using the multi-scale training method,the neural network randomly selects one of the three scales for training when learning the feature,so that the overall target size is obtained.More uniform,so that vehicles of all sizes can be better detected.Based on Udacity dataset validation,multi-scale training increased the average accuracy(AP)of detection by 2.4%.Aiming at the common convolutional neural network,the fixed position is used to sample the feature,which makes the network difficult to adapt to the geometric deformation of the object.The deformable position sensitive ROI pooling is proposed.Compared to the ROI pooling,an offset is added,which causes the offset to occur at the time of sampling,changing the original fixed sampling situation.On the Udacity dataset,the deformable network increases the AP by 0.5%.For the problem of target miss detection in complex environments,soft-NMS is introduced.When the coincidence between detection frames exceeds the threshold,the detection frame with lower score is not directly removed.By setting the weights appropriately,the score is reduced,so that the originally detected target will not become the target of missed detection due to the traditional non-maximum value suppression(NMS).The results of testing on the Udacity dataset show that soft-NMS increases AP by 1.5%.Based on the above test on the Udacity dataset,the experimental results show that the AP is increased by 4.3% compared to the original R-FCN detection framework.The detection framework adopted in this thesis has achieved good results in detecting small targets and targets in complex environments,and it also has certain robustness.In order to mark the target more accurately and determine whether each pixel is in the target area,Mask R-CNN is introduced to semantically segment the target.Screening the network containing the vehicle target from the MS COCO data set to train and test the network,trim the number of network layers,reduce the use of parameters,and improve the efficiency of the model.Aiming at the problem that the segment branch result is difficult to improve due to the inaccurate mask branch confidence in Mask RCNN,the Mask Io U branch is introduced,and the predicted mask is compared with the Io U of the original mask to calibrate the deviation from the actual.The AP of the segmentation effect was increased by 0.5%.