Research On Multi-target Tracking Algorithm For Intelligent Monitoring Syste

Intelligent surveillance systems have penetrated into all aspects of people’s lives to guarantee the safety of society.Multi-object tracking is a key technology in the image analysis stage of intelligent surveillance systems.It plays a pivotal role in the system which is the intelligent core of the system.Feature embedding and anchor free multiobject tracking framework are the technical key of intelligent surveillance system.They address the difficult issues such as variable monitoring routes,irregular observation angles and changing object appearance in the system.Despite these frameworks,there are still issues such as missing embedding features and motion information in data association,as well as insufficient noise control capability in both types.In this paper,the following two aspects of work are carried out to address the above problems.For the problem of missing embedding features and motion information in data association of feature-embedding multi-object tracking algorithm,this paper proposes a dual-factor driven multi-object tracking algorithm with embedding features and motion information.Firstly,we enhance the offset features embedding between frames by embedding feature information enhancement network.This improves the acquisition of object appearance representation information,expands the difference of features between frames,optimizes the information utilization of association matching.Secondly,we propose the GRU motion information complementation method to obtain the motion information of the object.This method is achieved by learning the temporal information between data points from the hidden state according to the regressed category scores and object locations.It is utilized to complement motion information during the association phase.The two are combined to complement the appearance representation and motion information in the data association phase,which jointly drive the improvement of the algorithm performance.The algorithm was experimentally validated on the KITTI dataset.The ablation experiments show that HOTA and MOTA improve 1.59% and 0.42% respectively compared to the baseline model,which verify the effectiveness and compatibility of our model.In addition,we compared our proposed algorithm framework with other mainstream algorithms,which verify the superiority of our algorithm.To address the problem of insufficient noise control capability in the anchor free multi-object tracking algorithm,this paper proposes a noise-control multi-object tracking algorithm.Firstly,the proposed heatmap feature prior denoiser obtains a priori information of the model from the jump connection.It eliminates redundant noise in the fusion process of heatmap and image features,which enhances the semantic representation capability of the fused features.Secondly,the proposed smoothing gain Kalman filtering algorithm combines the Gaussian function with the adaptive observation coefficient matrix to stabilize the abrupt noise of Kalman gain,which increases the motion information in data association.Finally,gradient boosting reconnection context mechanism is designed to achieve reconnection results consistent with the trend of real motion trajectories.It stabilizes the drifting trajectory noise.by adaptive interpolation of fragment trajectories through gradient-accelerated decision trees.The algorithm was experimentally validated on MOT16 and MOT17 datasets.The HOTA and MOTA are improved by 6.7% and 6.37% respectively in the ablation experiment compared with the baseline model,which verify the effectiveness and compatibility of our model.In addition,our paper also compares the proposed algorithm framework with other mainstream algorithms in the MOT Challenge to verify the competitiveness of our algorithms.