Research On Target Tracking Method For Indoor And Outdoor Scenes In Chemical Industry Park

Chemical industry parks are a common industrial structure in China and an important pillar of China’s economic development.Although it is a high-input and highoutput industry,it is also a high-energy-consuming and high-pollution industry.Therefore,how to promote industrial wisdom upgrading,scientific planning and strengthen the safety supervision of the park,so as to realize the green and intelligent development of enterprises,is the focus of the future construction of the park.As the country enters the digital era,massive information data plays a pivotal role in the security monitoring work of the chemical park,and the more intelligent digital chemical park also puts forward higher requirements for the prediction of risks in indoor and outdoor scenes.The combination of mature target tracking technology and indoor and outdoor scenes in chemical parks is of great significance to effectively reduce the security risk of parks and realize real-time monitoring and risk prediction of abnormal information in chemical parks.Based on the above background,this paper discusses the basic concepts of visual target tracking for indoor and outdoor scenes in chemical parks,as well as the mainstream algorithm research directions and research methods,and fully draws on some cutting-edge research results in the field of target tracking.We aim at solving the problems of tracking in complex scenes in chemical parks,the similarity of similar targets,and the time-consuming and labor-intensive manual annotation of large-scale video data in chemical parks,and the main research directions are how to effectively use fusion features for complete representation of target information and unsupervised visual tracking of a large number of unlabeled video sequences.The main research work and results of this paper contain the following aspects.(1)a target tracking method that fuses co-occurrence statistics and fhog features.The method adds a co-occurrence filtering tracking module to the discriminative scale space tracking algorithm.Firstly,the image size of each frame read is adjusted to the specified model size,then it enters the co-occurrence filter tracking module to do the operation with the original fhog gradient features of the position filter,and finally the feature matrix of the target image block based on the position information is obtained.Similarly,the target images at different scales are fed into the co-occurrence filter tracking module and the scale filter for the concatenation of features to obtain the feature matrix of the target image block based on the scale information,and then fed into the target tracking module on the basis of this feature.The algorithm combines the position information obtained from the position filter and the target scale obtained from the scale filter in the discriminative scale space tracking algorithm to continuously update the co-occurrence filter model,which can represent the global feature information of the target image according to different image details,thus improving the accuracy of target tracking.The experimental results show that the proposed improved algorithm not only outperforms other correlation filter-based algorithms in terms of distance accuracy and overlap accuracy,but also compares with the f DSST(fast Discriminative Scale Space Tracking)algorithm,which has the best accuracy,the algorithm in this paper is 7.1% higher than f DSST in terms of average overlap accuracy and 1.1% lower than f DSST in terms of average distance accuracy.The algorithm can track the target consistently and accurately,and has better robustness.(2)Unsupervised target tracking method based on deep convolution and selfattention mechanism.The current supervised learning target tracking method requires more manual labeling information,which is limited by the quantity and quality of manual labeling data and relies on a large amount of manual labeling experience,and cannot meet people’s expectations of artificial intelligence.On this basis,an unsupervised learning method based on a fused convolutional and attention mechanism is proposed.The convolutional layer has a strong inductive bias,which makes its network model generalize better and converge faster,and the attention layer can effectively utilize large-scale data due to the larger scale of its model.On this basis,combining the convolutional layer with the attention layer can improve the generalization performance and operation scale of the model.Therefore,a network model incorporating deep convolution and self-attention as the backbone network structure is constructed,and a combination of forward and backward tracking is used to achieve optimal tracking of the target through multi-frame verification,consistency loss function and other learning means.Experimental results show that this unsupervised tracker can achieve the benchmark accuracy of trackers under supervised learning,and the AUC metric on the OTB100 dataset improves 3.4% over the benchmark algorithm,and 4.6% over the benchmark algorithm on the OTB50 dataset.The effectiveness of the method for learning without labels is demonstrated.The method is superior when compared with the supervised method.The qualitative and quantitative analysis of the experimental results shows that modeling target information by unsupervised visual tracking and fused features can rely less on labeled data as well as effectively improve the expression of visual target features,thus closely integrating theoretical research with indoor and outdoor scenes in chemical parks and promoting practical applications of target tracking in security surveillance scenarios.