HIERARCHICAL MULTI-OBJECT TRACKING METHOD BASED ON SALIENCY DETECTION

摘要

#$%^&*AU2020100371A420200416.pdf#####ABSTRACT The present invention relates to a hierarchical multi-object tracking method based on saliency detection, including: creating a mixed data set based on an existing standard data set with manual annotations; constructing a saliency region detection sub-network for road traffic scenarios to generate a salient region and a non-salient region; using an object detection algorithm to detect an object in each frame; constructing a multi-object tracking network model that combines a convolutional network, an LSTM network, and a graph convolutional network, and tracking an object in the salient region; building a parallel KCF pool to perform single-object tracking on an object in the non-salient region; and combining and post-processing trajectories in the salient region and non-salient region, to generate an overall trajectory. According to the present invention, a bounding box in a salient region is generated by adding fast saliency detection, to input, detect, and track objects, thereby increasing a detection speed while maintaining detection accuracy. The present invention can reduce computation complexity and speed up multi-object tracking in real autonomous driving scenarios.DRAWINGS Collect traffic data, combine it with an existing multiobject tracking data set, and perform data augmentation Train a saliency sub-network, and use it to obtain a salient region and a non-salient region Input data to a YOLO v3 network to restrict object detection Yes Whether it is a salient region Input data to a multi-object tracking network to perform No multi-object tracking Enable multi-threading, and use a parallel KCF algorithm to perform single-object tracking on an object in the non-salient region Perform enhanced validation on tracking trajectories and bounding boxes in the salient region and non-salient region End FIG. 1