Research On Image Detection And Tracking Of Human And Vehicle

Visual detection and tracking of human and vehicle are the basic content and research hotspots in the field of computer vision.They are widely used in video surveillance,intelligent vehicle driving,robot navigation and human-computer interaction.These applications can greatly accelerate informationization and automation processes of the logistics industry.Based on computer vision technologies and deep learning methods,this thesis studies the human vehicle detection and tracking algorithm.Two tracking algorithms are designed based on the twinning network and related filtering principles,which are end-to-end target tracking method and long-term tracking method combined with re-detection.The main research work of this paper is as follows:(1)Aiming at the auto-driving picture data with high resolution and small target,based on MaskRCNN,a multi-scale cutting training and multi-scale testing strategy are proposed.In the training phase,multi-scale cropping of the original image,generate a large number of training samples,retain and enhance the details of the small target;in the test phase,the original image is multi-scale scaled,and the results are integrated.It effectively solves the problem that the large resolution image accounts for too high memory and improves the effect of the model on small target recognition.Using the Apolloscape dataset for evaluation,the algorithm performed superiorly and achieved the fourth place in the CVPR 2018 WAD(Workshop on Autonomous Driving)Challenge.(2)LR-AFNet is proposed,an end-to-end tracker for target location(via the Siamese network)and bounding box regression,to better locate and estimate scales,and using attention-based multi features fusion method to extract more discriminative features.Firstly,the feature extraction backbone network is redesigned to improve the location accuracy.Secondly,in order to obtain more effective features,deep and shallow features are fused with an attention mechanism,which adaptively learn the fusion weights.Finally,the bounding box regression branch is added to the Siamese similarity network,forming an end-to-end trained framework.Experiments proved that our proposed method achieves competitive performance on several benchmark datasets,compared with the state-of-the-art.Compared with the long-term tracker in the third part,LR-AFNet is aimed at the general tracking scenario,and requires a large amount of data and computing resources for large-scale training in advance,but can get results end-to-end through one network,simplifying the process.(3)In order to accomplish the long-term visual tracking task in complex scenes and solve the problems of scale change,occlusion,target disappearance of the camera view and tracking failure,this paper proposes a long-term scale adaptive tracker based on correlation filter and combined with instance matching re-detector.Our method is composed of tracker,confidence module and re-detector.Firstly,in the correlation filter tracker,the features of multi-scale search areas are extracted by feature pyramid network and adaptive deep-shallow features fusion method,to improve the scale adaptability and robustness of the tracker.Then,a Prediction Quality Measure index that simultaneously indicates the robustness and accuracy of the tracking response map is used to determine whether the tracking failed.Finally,when the tracking fails,the re-detector is applied in expanded search area,and an instance matching method is designed to remove the distractors,determining the target.This algorithm is applied to challenging benchmark OTB2015,VOT2016 and VOT-LongTerm2018,which contain long-term video sequences.Compared with state-of-the-art trackers,the algorithm improves the tracking performance and shows robustness in long-term visual tracking of complex scenes.Compared with the end-to-end tracker in the second part,the algorithm has many parameters and the process is complex,but the training requires less resources,and it solves the problem of target scale change,appearance change,tracking disappearance and occlusion in long-term scenarios.