Multi-object Tracking Algorithm Based On Automatic Multi-cue Fusion

Multi-object tracking is an important sub-topic in computer vision tasks,in smart security,automated driving,automated navigation,missile guidance,group analysis,etc.are of great importance.A lot of work has long been devoted to improving the accuracy of multi-objective tracking and trying to build a real-time multi-objective tracking system.Along with the vigorous development of object detection technology,the development of detection-based multi-objective tracking algorithms is also extremely rapid.Existing multi-object tracking algorithms are based on target expression,cue fusion,and allocation association,and are constantly improving their effectiveness.However,the existing algorithms still exist in different degrees of slow running,feature is not robust,can not handle the detector error and other problems,this paper wants to overcome in terms of strengthening the cue expression and data association.In terms of enhanced cue expression,the detection algorithm is proposed in parallel with the extraction of features to solve the timeliness problem and make the multi-object tracking algorithm Tracking algorithms become possible in real time.A multi-cue strategy combining global and local features is proposed to solve the problem of detector errors and potential occlusion.In terms of data correlation,an adaptive multi-cue fusion strategy based on contextual content is proposed,differentially assigning fusion to each target track coefficients,specifying the following research methods.First,we propose a parallel multi-object tracking method based on feature fusion.The traditional feature map fusion strategy has the problem that the high-level information conveyed during the fusion process is weakened,and we propose cross-layer downward cascading to solve the semantic loss case by cascading the high-level feature map obtained by Res Net-50 with all the low-level feature maps.The semantic information is more enhanced on the feature map of each layer.In order to improve the speed of the algorithm,we propose a parallel strategy of extracting features from the target while detecting it,so as to get rid of the problem of "detect first,extract later".features" model,which we also compare on the validation set.Second,a locally-aware multi-object tracking algorithm is proposed.To solve the problem of inaccurate detection algorithms and misaligned global features.We propose a local alignment strategy based on locally-aware classification,which focuses on overcoming the challenges of half-blocking and poor detector performance The target adopts the use of visible local features for matching,which is more in line with human thinking and is compared with the algorithm in apublicly available evaluation platform.Third,we propose a context-based adaptive multi-cue fusion tracking algorithm.In order to effectively combine the partial cues,we propose a context-based adaptive fusion cue strategy,which considers the surrounding area of different targets.The properties of other targets in the neighborhood are optimally combined for multiple cues using differential fusion coefficients.Finally,this strategy is added to several manual parameter setting experiments and adds discussion.