Deep Multi-Shot Network for modelling Appearance Similarity in Multi-Person Tracking applications

The automatization of Multi-Object Tracking becomes a demanding task in real\nunconstrained scenarios, where the algorithms have to deal with crowds,\ncrossing people, occlusions, disappearances and the presence of visually\nsimilar individuals. In those circumstances, the data association between the\nincoming detections and their corresponding identities could miss some tracks\nor produce identity switches. In order to reduce these tracking errors, and\neven their propagation in further frames, this article presents a Deep\nMulti-Shot neural model for measuring the Degree of Appearance Similarity\n(MS-DoAS) between person observations. This model provides temporal consistency\nto the individuals' appearance representation, and provides an affinity metric\nto perform frame-by-frame data association, allowing online tracking. The model\nhas been deliberately trained to be able to manage the presence of previous\nidentity switches and missed observations in the handled tracks. With that\npurpose, a novel data generation tool has been designed to create training\ntracklets that simulate such situations. The model has demonstrated a high\ncapacity to discern when a new observation corresponds to a certain track,\nachieving a classification accuracy of 97\\% in a hard test that simulates\ntracks with previous mistakes. Moreover, the tracking efficiency of the model\nin a Surveillance application has been demonstrated by integrating that into\nthe frame-by-frame association of a Tracking-by-Detection algorithm.\n