Visual Multiple Object Tracking In Complex Background

Multiple object tracking is an important research field in artificial intelligence.It has great significance in other computer vision tasks such as intelligent navigation,intelligent transportation,security monitoring,and abnormal behavior analysis.Although a great achievement of tracking has been obtained in recent years,the uncertainty of the numbers of interest objects,the complexity of background,the appearance similiarity between objects and background and strong occlusion,fast and robust tracking of multiple objects in real applications is still highly challenging.This work focuses on graph based tracklet association and object appearance modeling using a deep convolutional neureal network.The major contributions and results are described as follows.We propose a semi-online MOT method using online discriminative appearance learning and tracklet association with a sliding window.We connect similar detections in neighboring frames with a temporal window,and improve the performance of appearance feature by online discriminative appearance learning.Then,tracklet association is performed by minimizing a subgraph decomposition cost.Occlusions and missing detections are recovered after tracklet stitching.Our method has been tested on two public datasets.Experimental results have demonstrated the significant performance improvement of our method.Specifically,the proposed method is respectively improved by 8.31% and 12.38% in term of MOTA and MOTP as compared to the baseline.Detections and appearance feature play an important role in MOT,but traditional hand-drafted appearance feature cannot fully distinguish similar targets.This paper proposes a pragmatic method to extract more robust appearance feature using a convolutional neural network.We adopt the detection results produced by a state-of-the-art detector and extract the appearance feature by a convolutional neural network.In our method,the motion state is first estimated by Kalman filtering.The appearance feature is further combined with the motion state to estimate the affinity scores between different detections and associated trajectories.Finally,detections are assigned to their corresponding trajectories using the Hungarian algorithm.Our method has been tested on the MOT16 dataset,the evaluation results demonstrated that combining motion state with the appearance information extracted by a neural network is efficient for continuous and robust tracking of multiple objects.