Methods, devices and computer programs for distance metric generation, error detection and correction in trajectories for mono-camera tracking

摘要

A surveillance system analysis method comprises determining a distance metric for comparing target images belonging to trajectories obtained by tracking potential image sequence targets obtained from one camera, each trajectory being associated with/comprising successive sequence target images. Each target image comprises an image portion corresponding to a potential target; trajectories are defined by successive target images associated with the potential target. The method comprises: extracting a set of features from each target image; generating plural chunks (image sets): a predefined number of consecutive target images in trajectories; generating plural positive pairs, belonging to a same chunk; generating plural negative pairs (pairs belonging to chunks being part of two different trajectories, chunks co-existing at the same time); and training a classifier with the positive and negative pairs and the set of extracted features to define the distance metric. A method is also disclosed wherein: initial sequence trajectories are obtained, from mono-camera tracker(s), defined by a succession of bounding boxes and an associated same target identification (ID) tag; for each pair of plural pairs of bounding boxes, similarities between bounding box content are measured; initial trajectories are split based on similarities; and trajectories are generated, linking split trajectory portions.