Transformer-based Multi-pedestrian Tracking

Deep learning technology is widely used in computer vision field.Researchers began to apply deep learning technology and methods to solve the multi-pedestrain tracking tasks.At present,the paradigm tracking-by-detection widely used in tracking technology,and the detection effect affects the tracking effect.Multi-pedestrain tracking task can be divided into target detection and re-identification.On the basis of good detection effect,re-identification is the key to determine whether it can track stably.According to the tracking-by-detection paradigm,the multi-pedestrian tracking task is divided into two sub tasks: pedestrian detection and re-identification.In order to get better pedestrian detection and re-identification effect,this paper combines the timing information and transformer model.This paper contributes the following ideas:Modify transformer model's calculation method of the attention mechanism,so that different frames information can interact in the same batch,and previous batch's content can be used in current batch.Previous Encoder module's output is being the input of current,and this way is used to realize the utilization of timing information.This paper also puts forward a new pos-encodings which expresses the location information in a more intuitive way and strengthens the role of spatial information.In the pedestrian re-identification,the classification network is used to associate pedestrians,which improves the stability of pedestrian identification in the process of pedestrian tracking.In summary,a multi-pedestrian tracking model using spatio-temporal information is realized to improve the stability and comprehensive tracking performance of pedestrian tracking signs.Based on the above work,the tracking model based on transformer makes a good performance in MOT Challenge datasets and evaluation metrics MOTA,reached 63.3 on MOT15 datasets and 70.5 on MOT17 datasets,having good tracking effect and on both datasets,the identification switch evaluation IDSW,reached a great effect,342 on MOT15 and 2321 on MOT17.Both evaluation metrics of multi-pedestrain tracing higher than comparable tracking model.