Convolutional Neural Network-based Simple Online Multiple Object Tracking

Online multiple object tracking (MOT) is difficult when multiple objects have similar appearance or under occlusion. Detections and appearance feature play an important role in MOT, but the distinguishing ability of traditional hand-draft appearance feature is in sufficient. This paper explores a pragmatic method to extract more robust appearance feature using convolutional neural network. We adopt the detection results processing by the state of art detector and extract the appearance feature by a convolutional neural network. In our method, the motion state is firstly estimated by Kalman filter. Then 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 by Hungarian algorithm. Our method has been test on standard MOT dataset MOT16, the evaluation results confirmed that combine motion state with the appearance information extracted neural network is efficient for tracking multiple objects continually and stably.