Research On Highly Robust Target Tracking Algorithm In Complex Scenes

Object tracking is an essential and challenging research direction in the field of computer vision.In recent years,with the rapid development of computer hardware and the emergence of algorithms with excellent performance,object tracking technology has gained lots of traction in commercial,medical and military domains.However,due to the existence of interference factors such as occlusion of the object,deformation of the object,illumination variaation,and the presence of similar objects in the background,the current algorithms perform sub-optimally in complex scenes.Therefore,in order to address the problem of highly robust object tracking in complex scenarios,this paper starts with two aspects of the tracking algorithm based on the correlation filtering framework and the end-to-end tracking algorithm based on the deep convolutional neural network,and conducts the following research work:1.A multi-feature fusion object tracking algorithm based on kernel correlation filtering framework is proposed,which simultaneously extracts the HOG features,CN color features,and Gray features of the image for tracking.For the multi-feature fusion strategy,adaptive fusion based on the confidence of the response map is selected.At the same time,an improved scale adaptation mechanism is introduced to adapt to the scale change of the target during the tracking process.Further,based on the confidence of the response map and the similarity of adjacent frames,a mechanism for dynamically adjusting the learning rate is set to reduce the effect of the non-ideal tracking result on the model update.It is shown by experiments that the algorithm has a tracking accuracy rate under the 20-pixel threshold and an average tracking success rate of 0.796 and 0.596 on the OTB100 dataset,which indicates that it has high tracking robustness.Comparative experiments also confirmed the effectiveness of the three adaptive adjustment mechanisms set in the algorithm.2.An object tracking algorithm based on the dual Siamese network is proposed.The algorithm uses three modified VGG-16 networks to extract features from the current frame search area,the first frame template area,and the previous frame output area.Then,two similarity score maps are calculated and obtained.And then the score maps fusion is performed based on the adaptive fusion strategy,and the object positioning is performed based on the fusion score map.During training,the triplet loss is used for model training to obtain better model parameters.At the same time,a mechanism for dynamically setting the third input according to the confidence of the tracking result is added to improve the robustness of the model.The algorithm does not update the model during the tracking process to improve the real-time performance and prevent the model from being polluted by the wrong tracking result.The tracking accuracy rate on the 20-pixel threshold and average tracking success rate of the algorithm on the OTB100 dataset are 0.860 and 0.641,respectively,indicating that the algorithm performs well in complex environments and has high tracking robustness and superiority.3.Based on the two tracking algorithms proposed in this paper,a visual tracking system in complex scenes is designed and implemented.The system uses the SHANYI-30 as the tracking platform,Hikvision DS-2CD2820 F camera as the visual sensor,and the edge computing devices in vehicle is used to execute the tracking algorithms.Corresponding tracking software is designed based on the tracking algorithms.The experimental results in real scenarios show that the two algorithms proposed can well overcome the interference factors in the scenes and perform highly robust real-time object tracking in complex scenes.