Real time multiple object tracking using deep features and localization information

In this paper we propose a tracking by detection method using a dissimilarity measure calculated based on the location and the appearance information of the object. These dissimilarity values are used in Hungarian Algorithm [1] in the data association step for track identity assignment. We make use of YOLO [2] deep learning based object detector in the detection step from camera image feed. Location measure is calculated using the predicted object location and bounding box, while the appearance measure is from the last feature layer from the detection network. Main focus in this work is to propose a tracking framework that can be used in real time automated vehicle guiding applications, by striking a balance between computational complexity and tracking accuracy. Therefore, we make use of the deep features available from detection framework rather than calculating a new appearance measure during the tracking step. The method proposed is very efficient and enables to achieve speeds up to 500+ frames per second (fps) in KITTI [3] tracking benchmark while achieving state-of-the-art results.

IEEE 15th International Conference on Control and Automation