| Multiple-Object-Tracking(MOT)can realize real-time tracking by detecting the identity information of multiple targets,which is a key technology in the field of computer vision.MOT has played an excellent role in the fields of artificial intelligence such as military,security and autopilot.However,the current MOT algorithm cannot accurately identify and track targets in scenes with dense target distribution or more occlusion,and the tracking results contain many false results.In addition,due to the widespread application of the detection-based MOT paradigm,the module compatibility problem caused by the paradigm has virtually restricted the performance of the multi-target tracking model.Finally,the pursuit of performance of convolution neural network undoubtedly leads the development of MOT model to the construction of large model,which cannot guarantee the accuracy and reduce the speed of model reasoning.To solve these problems,this paper proposes the following solutions based on attention mechanism and neural network architecture search algorithm:(1)For the occlusion problem in MOT scenarios,this paper proposes an adaptive spatial attention module and an adaptive temporal attention module based on an adaptive attention distribution mechanism,and uses a feature extraction network and a bounding box regression network to refine Detect and predict the position,and based on the intersection and union ratio of the bounding box and the predicted box,the occluded target and the unoccluded target are processed in different ways during the tracking process,effectively dealing with the severely occluded target and alleviating identity conversion situation,so as to achieve stable tracking results.(2)A one-step MOT network is proposed,a global feature aggregation network is designed for the fusion of different levels of features,and based on the traditional target detection network,the functional head is extended into a composite task reasoning branch,At the same time,the reasoning of the target detection reasoning branch and the re-identification reasoning branch is completed.The global feature aggregation network adds cross-layer fusion of deep features and shallow features based on ResNet,which effectively enhances the mutual compatibility between modules in multi-target tracking tasks.(3)Aiming at the high inference delay problem of the multi-target tracking model,a single-path,equal-probability neural network architecture search algorithm is proposed.Based on the single-step search framework,equal-probability fair training is performed on each subnetwork path,which can Effectively improve the feature expression ability of the sub-network,and then relatively enhance the performance of the lightweight target tracking model.In addition,the constrained evolutionary algorithm is used to search the sub-network in the super-network,which effectively reduces the search time,improves the search efficiency,and increases the probability of finding an excellent sub-network.This paper conducts experiments on multiple industry-recognized and widely used public datasets and compares it with other multi-target tracking algorithms with good performance.Experiments prove that the method proposed in this paper can effectively alleviate the above problems and improve the performance of MOT algorithms. |
