Abstract
This thesis presents a novel system for high performance multi-vehicle tracking. The main concept of our system is tracking by detection, or to say, detection based tracking. We consider the whole tracking problem as a combination of per-frame object detection, detection result regression, embedding metric of detection result, and association of final detection result. The core technique in implementation is convolution sharing. We share the same rich features between convolutional neural network (CNN) models that are designed for different parts of our system. By integrating the concept and the core technique mentioned above, with a properly scheduled training process for specific CNN models, our system achieves a good performance when evaluating on the object tracking datasets of KITTI, a challenging real-world computer vision benchmark suite. Moreover, even though our tracking system is built on multi-stage and multi-CNN models, it can run efficiently in nearly real time.