| In open environments such as forests and prairies,the lighting conditions are complex and changeable.Meanwhile,the obstructions are often overgrown.Those circumstances significantly increase the difficulty of wildlife protection with object tracking technology.In addition,actions including the illegal hunting prevention and the wildlife rescue often need to be carried out at night and in the early morning when the light conditions are poor.The object tracking platform is required to have all-weather target tracking capability to better complete the wildlife protection tasks utilizing the object tracking technology.If the all-weather target tracking is required to be achieved,it is not enough to use only the RGB images collected by the visible light camera.Because the visible light camera cannot make clear images at night or in extreme visual environments such as heavy fog,rain and snow,while the thermal infrared camera can obtain better images is in the above extreme visual environments.As a result,this paper has conducted the research in these three aspects: visible light object tracking method,thermal infrared object tracking method,and dual-modality fusion of visual light and thermal infrared object tracking method.We also propose a wildlife protection technical scheme based on the UAV platform to verify the feasibility and advancement of the three object tracking methods proposed in this paper.The main research contents of this paper are as follows:(1)To solve the problem of low object search efficiency of current methods in the longterm object tracking task in the visual light modality,a two-stage object tracking method based on the dual attention mechanism is proposed.In order to improve the deep feature extraction ability of the backbone network and reduce network parameters,the proposed method uses Efficientnet-B0 as the backbone network for feature extraction.The accurate estimation of the target bounding box is performed by maximizing the DIo U for bounding box regression.The dual attention based on channel and space is added to the target classification module to further improve the discrimination ability of the tracker to the target and background.The most important innovation of the object tracking framework is the two-stage search method and target screening module.In the first stage of search,the sliding window is used to search the area with high possibility of target existence.In the second stage of search,the screening module is used in the region with low possibility of target existence,and the top K regions with the highest possibility of target existence are screened and sent to the inference branch for the next step of inference,while the other remaining regions are directly discarded.A large number of experiments demonstrate that this method has a good balance between three performance metrics of accuracy,robustness and real-time performance and performs well especially when dealing with the challenge of target re-entering the field of view after the target appears in the long-term object tracking task.On the La SOT dataset rich in typical challenges of long-term object tracking,the success rate index of this method can reach 0.523,and the precision index can reach 0.518.(2)To solve the problem of insufficient ability of the current methods to resist the occlusion and interference between similar objects in the object tracking task of thermal infrared modality,a object tracking method based on efficient global information perception is proposed.In order to efficiently obtain the global semantic information of the image,this paper utilizes the Transformer structure for feature extraction and fusion.In the feature extraction process,the Focal Transformer structure,which is highly similar to the human attention mechanism,is used to improve the efficiency of information remote modeling.In the feature fusion process,the relative position encoding is supplemented to the standard Transformer structure,which allows the model to continuously consider the influence of positional relationships during the learning process and generalize to the capture of different position information contained for different input sequences,so that the Transformer structure can more efficiently model the semantic information contained in the image.To further improve the tracking accuracy and robustness,the heterogeneous bi-prediction heads are utilized in the object prediction process.The fullyconnected sub-network is responsible for the classification prediction of the foreground and background,and the convolutional sub-network is responsible for the regression prediction of the object bounding box.In order to alleviate the contradiction between the huge demand for training data of the Transformer model and the insufficient scale of the thermal infrared object tracking dataset,the La SOT-TIR dataset is generated with the generative adversarial model for network training.Through systematic experiments,it is found that the attention mechanism of Transformer plays an obvious role in improving the performance of thermal infrared tracker.The proposed method performs particularly well when dealing with the challenge of severe occlusion and similar interference,and the speed can reach 90 FPS.(3)To solve the problem of the low data fusion efficiency of the current object tracking method of visible light and thermal infrared dual-modality data fusion,the RGB-T object tracking framework CEDi MP based on channel exchange for dual-modality data fusion is proposed.The CEDi MP utilizes the scale factor of batch normalization(BN)as the importance measurement of each corresponding channel during feature fusion,and replaces the channel that has the scale factor close to zero to the current modality with the value of the other modality,which allows for the dynamic exchange of information between the sub-network in the visible light and the thermal infrared modality.A large number of experiments prove that the feature fusion methods based on channel exchange are simple and efficient,which can not only improve the accuracy and robustness of the RGB-T object tracking framework,but also improve the speed of the tracker.Since CEDi MP utilizes Di MP as the basic tracker,the ability of CEDi MP framework to distinguish the target background is significantly better than the existing RGB-T tracking framework under the strong power of Di MP’s online learning ability.In addition,using the artificial synthetic dataset La SOT-RGBT for neural network training not only improves the ability to deal with the typical challenges of long-term object tracking,but also significantly improves the generalization performance of the tracker.CEDi MP can be faster than 20 FPS,and double mf Di MP.(4)To solve the problem that the current wildlife protection technology cannot keep up with the needs of wildlife protection,an efficient and feasible wildlife protection technical scheme is developed based on the UAV platform and the proposed object tracking method in Chapter 6.The scheme involves technologies containing the UAV trajectory planning,object detection,active object tracking and remote wireless communication.Experiments on two specific tasks including the wildlife search and rescue as well as the illegal hunting behavior prevention demonstrate the feasibility of the scheme.The effectiveness and advancement of the object tracking method proposed in this paper is also proved.In Section 6.3,the comparative experiments are conducted between the proposed method and the KCF object tracking method,which is the most widely used UAV platform currently.Through experiments,it is found that the object tracking method proposed in this paper is better than the KCF method in both accuracy and robustness... |
PDF Full Text Request