Research On Multi-object Tracking Related Algorithms In Complex Scenarios Based On Multi-feature Fusion

With the rapid development of science and technology,human society has entered the era of artificial intelligence.The explosive growth of videos data generated by numerous intelligent device has attracted widespread attention to computer vision technology.Mul-tiple target tracking technology in video is an important research topic in the field of com-puter vision.High quality multiple target tracking algorithms are the foundation for solving more advanced visual analysis and decision-making tasks,including video surveillance,au-tonomous driving,posture estimation,action recognition,behavior analysis,and anomaly warning.The main factor affecting tracking performance is the model representation and algorithm implementation of the appearance features of the target during tracking.How-ever,due to the existence of multiple complex real-world factors,such as occlusion,target deformation,background changes,etc.,learning an accurate appearance model becomes ex-tremely complex.Based on deep learning and integrating multiple features of targets,this thesis conducted research on multi-objective tracking related algorithms in complex scenes.The main research content and innovative work are as follows:（1）A multiple target tracking method based on particle filter-wavelet packet trans-form-BP neural network is proposed to address the issue of particle filter based method not working properly when partial or complete occlusion of the face occurs during multiple per-son face tracking.This method mainly consists of three steps.The first step is to use a face detection algorithm to detect the faces in the first frame of the image,perform wavelet packet transform,and extract color features.The second step is to apply feature fusion to update particle weights,and use the proposed BP neural network to track occluded faces.And the third step is to use particle filtering to track faces.By applying wavelet packet coefficients and fusing multiple features such as traditional color features,facial features are effectively described.Wavelet packet decomposition is used to generate some frequency coefficients of the image,and the high-frequency and low-frequency coefficients of the reconstructed signal are used to improve facial tracking performance.To improve the robustness of the algorithm for occluded faces,a BP neural network is used.The performance of the proposed algorithm was evaluated through numerical experiments,verifying its effectiveness.（2）A tracking device GLITrack,which integrates multiple optimization techniques with strong real-time performance and high robustness,is designed to address issues such as target occlusion,deformation,and real-time requirements during multiple target tracking in complex scenes.This tracker obtains global information of trajectories by establishing a global link model that only uses spatio-temporal features of trajectories without appearance features.It uses a method based on camera motion compensation to reduce issues such as target mismatch and body identification switch activation caused by camera motion jitter,thereby achieving accurate positioning of targets in moving scenes;The Gaussian process regression method is used to smooth interpolate the data and compensate the missing data,which avoids the shortcomings of linear interpolation in ignoring the motion information and limiting the accuracy of the interpolation position.The noise scale adaptive Kalman filter algorithm is used to denoise,which can adaptively adjust the noise scale according to the detection quality.The evaluation of the designed tracker GLITrack on multiple target tracking datasets shows that the tracker effectively improves the overall performance of multiple target tracking.（3）A pixel level kernel prediction network model is proposed to address the diffi-culty of tracking multiple targets in harsh rain and fog environments.It can effectively eliminate rain streaks,reconstruct high-quality backgrounds,and achieve robust image de-noising.Existing precipitation models usually require specific assumptions,but in practical applications,they are difficult to cover multiple scenarios and require complex optimization or gradual improvement.Therefore,a kernel prediction network based on Unet++was con-structed to estimate pixel level kernels from rainwater images,achieving effective removal of image raindrops.In order to improve the rain removal effect of images,a loss function composed of structural similarity index loss,edge loss and L₁loss was used,and data opti-mization methods are combined to minimize the error between synthetic data and real data,and improve the performance of processing real rain images.The performance of the pro-posed method was evaluated through experiments,and the experimental results show that this algorithm has a higher peak signal-to-noise ratio than traditional methods.（4）A multiple target tracking method MSDATrack based on multi-scale deformable attention mechanism is proposed to address the issues of target tracking loss or inaccurate tracking in complex environments（such as target occlusion and adverse weather）,which can better extract target features in multi target tracking.Unlike traditional attention mecha-nisms,multi-scale deformable attention utilizes normalized position encoding to effectively address the issue of misaligned target features.A target tracking model based on multi-scale deformable attention mechanism has been designed to solve the problem of tracking loss caused by scale changes and partial occlusion of tracked objects in multiple target track-ing.According to the model,the corresponding loss function is given.Through iterative optimization,multiple target tracking in complex environment is realized.The effective-ness of the proposed method was verified through experimental methods,and the overall performance improvement of the proposed method was also demonstrated in datasets such as MOT17 and MOT20.