Discrete-Continuous Optimization for Multi-Target Tracking (CVPR 2012)


In this work, we formulate the task of tracking multiple targets as minimization of a single discrete-continuous energy. We follow the well established
tracking-by-detection approach where we first obtain a set of
independent target hypotheses
by an offline trained object detector. Then, given the unlabeled detections and a set of initial trajectory hypotheses, an alternating optimization scheme performs data association on the detections and trajectory estimation in continuous space by minimizing a single, discrete-continuous objective. Here, we show the
optimization process in more detail, where you see for each iteration
the discrete relabeling step, the continuous trajectory estimation
of the currently active models, the set of all possible trajectory hypotheses and the individual componenents of the energy. We now show the final tracking results
obtained by our method. Plese note that we show the raw tracker output
without any post-processing steps. Our proposed discrete-continuous optimization
is able to correctly perform data association while at the same time it produces smooth,
natural looking trajectories.

13 Replies to “Discrete-Continuous Optimization for Multi-Target Tracking (CVPR 2012)”

  1. no kxxxxa , i was not thinking about detection but about tracking, what if the camera is moving ? imagine a big translation of 100pixels between two frames. would it work ?

  2. Hello there Frank here, im currently looking to adding a Vision technique to my humanoid application currently i have a C# and C++ A.I program that takes advantage of the speech Recognition as well as Speech Synthesizer so i can be listened and it can reply by all this is just auditory systems working together and would like to learn about adding some sort of OpenCV Computational Visual system just wondering if you guys have anything to help me in my quest as i try to map its auditory input to a visual input in hope to create a 3D virtual environment mapped to the our analog world, many thanks in advance for your help and dedication in the field of study at hand.

    any c c++ or c# code  for this will be what ill be looking for  thanks,  did check the link but is not working

  3. Hi! May I know how you made the data visualization at 0:42 – 1:10, please?
    The 'path graphs' look neat on object tracking data outputs. 🙂

Leave a Reply

Your email address will not be published. Required fields are marked *