Gaussian mixed-trajectory prediction model for multi-target tracking
To address the challenge of increased difficulty in tracking object associations due to occlusions among targets in videos,a Gaussian Mixture Trajectory Prediction model is proposed to enhance the association performance under occlusion scenarios.This model employs a Transformer architecture to capture longer temporal dependencies from the historical motion information of objects,thereby accurately predicting the future trajectory distribution of tracked objects.Additionally,the model sets the output as a Gaussian mixture distribution,which combines multiple Gaussian distributions to define the future trajectory distribution of the tracked objects.This enables the utilization of trajectory distributions to calculate the spatial similarity between detection targets and tracked objects.Experimental results demonstrate that this model improves tracking performance,achieving a HOTA score of 56.3%on the MOT17 dataset.
multi-target trackingTransformermotion predictionGaussian mixture model
国家科技期刊平台
NSTL
万方数据
