Research On Online Multi-object Tracking Algorithm Based On Detection

Multi-object detection and tracking method is an important part of visual field,which has been widely used in intelligent driving,public security and other fields.Due to the complex and changeable realistic environment,the irregularity of moving objects and the unpredictability of object quantity and other internal and external interference factors,the existing multi-object tracking algorithms still have great shortcomings in accuracy and robustness when applied to practical problems.With the support of key science and technology projects of fujian province(2017H6009,2018H0018),this paper proposes online multi-object tracking methods based on detection,aiming at complex situations such as occlusion,scale change and unfixed number of objects in the multi-object tracking process,and improves the accuracy of object detection and tracking algorithms respectively on this basis.The main research results of this paper are as follows:Firstly,an object detection method based on deep learning framework multiple optimization strategy is proposed.Aiming at the problems of object detection in complex scenes such as false detection and missed detection,an object detection method based on deep residual network is proposed.The algorithm first forms five multi-scale prediction networks by adjusting the depth residual network,then introduces the second-order term to the residual connection in the process of deep fusion of feature graph to reduce the loss,and finally adopts the data amplification strategy in the training stage of the model.Experiments on the PASCAL VOC dataset show that the accuracy of the proposed model is 2.1% higher than that of the original model.Secondly,a correlation filtering method with scale ratio and appearance model adaptation is proposed.Aiming at the problem that the existing correlation filtering algorithm leads to tracking drift in the case of object scale change and occlusion,the improvement methods are proposed respectively.The algorithm introduces a scale filter and an adaptive model update strategy in the correlation filtering method.In the scale filter,the range of the scale pool is adjusted according to the change of the object scale of the front and rear frames,so that the scale ratio is adaptive,and the adaptive update strategy of the appearance model is added to the algorithm to enhance the discriminative power of the model.The experimental results show that the improved correlation filtering method improves the overlap success rate and the center distance accuracy by 10.1% and 1.9% respectively,and the algorithm showed stronger robustness.Thirdly,an online multiple object tracking algorithm for markov decision making based on correlation filtering optimization is proposed.Aiming at the problem that the existing multi-object tracking algorithm reduces the accuracy of multi-object tracking under frequent occlusion and interaction,the algorithm incorporates improved correlation filtering in Markov decision process,using correlation of scale and appearance model adaptive.The filter tracking is used as a tracking module to track multiple objects,and the history information of the tracking object,the maximum response value of the correlation filter,and the average peak correlation energy are combined as constraints to determine whether to continue tracking or to lose state,effectively reducing the error transfer of each object state.The experimental results in the MOT2015 dataset show that the proposed algorithm improves the multi-object tracking accuracy index by 2.2% and the ID error conversion number by 7.8%.In summary,because the object detection and tracking algorithms still have some shortcomings,this paper uses targeted optimization strategies to effectively improve the accuracy of the algorithm.Compared with other algorithms on the corresponding dataset,the experimental results show that the improved algorithm shows better robustness in various complex scenarios.