| The object tracking technology is a crucial aspect in the field of computer vision,achieved by recognizing and locating targets in video or image sequences,enabling continuous tracking of targets in time and space.Object tracking technology is widely used in video surveillance,autonomous driving,medical image analysis,robot vision,and other fields,analyzing and predicting target behavior through motion trajectory,speed,acceleration,and other characteristics.Siamese network-based object tracking technology has made significant progress,but still faces many challenges in complex scenarios,such as severe occlusion,illumination changes,background interference,and scale changes.In response to these issues,the main research content of this thesis is as follows:(1)To address the problems of target background interference and high algorithm complexity,a Siamese network tracking method based on the gating mechanism is proposed.Siam-FC is a classic Siamese network tracking model that uses Alex Net to extract features,but its shallow depth limits its ability to recognize deep features.Therefore,this thesis introduces an improved lightweight Shuffle Net V2 as the feature extraction network,achieving a balance between efficiency and accuracy.In addition,a temporal context gating network layer is added to adaptively update the temporal gating weights based on the historical tracking results,controlling the importance of each feature and enabling the tracking model to adapt to target changes.In relevant experiments with the OTB2015 tracking dataset,the algorithm’s precision and success rate reached 0.907 and 0.687,respectively,indicating strong anti-interference capability.(2)To address the problem of targets being occluded for a long time or severely,a object tracking method based on a dual-template Siamese network is proposed.This thesis introduces a dual-template Siamese network composed of the previous frame template and the first frame template,enabling adaptive template updates and effectively solving the problem of target loss caused by long-term or severe occlusion.Furthermore,an improved channel attention mechanism is introduced to adjust weights according to channel feature effectiveness,increasing the utilization rate of effective information.A spatial attention mechanism is also introduced to reduce the search range of the target during tracking,improving the tracking model’s speed and efficiency.In relevant experiments with the OTB2015 tracking dataset,the algorithm’s precision and success rate reached 0.912 and 0.707,respectively,indicating that the algorithm helps improve the robustness of object tracking in occlusion scenarios.(3)To address the problem of insufficient lighting and low image contrast in underground scenes,an improved Retinex method is proposed.This method first converts the RGB color space to the HSV color space to optimize the brightness component;then,it uses a block processing method to set brightness weights for each block,addressing the issue of detail loss caused by overexposure in highlighted areas.Additionally,an adaptive contrast enhancement method is adopted to increase image contrast.Finally,the HSV color space is converted back to the RGB image,completing the image enhancement process.The experimental results show that the image quality is significantly improved,helping to increase tracking accuracy.(4)Based on the aforementioned algorithms,an object tracking system is designed in this thesis.The system is designed according to scene requirements and implements registration/login,online/offline tracking,and other functions.The system has a user-friendly interface and simple operation,meeting the needs of an object tracking system.The tracking methods proposed in this thesis can achieve good accuracy and real-time performance in complex scenarios,with strong robustness and anti-interference capabilities.They improve the performance of existing research methods in complex scenarios and are expected to make a positive contribution to subsequent research. |
PDF Full Text Request