MOTChallenge. A Benchmark for Single-camera multi object tracking

In the recent past, the computer vision community has relied on several centralized benchmarks for performance evaluation of numerous tasks including object detection, pedestrian detection, 3D reconstruction, optical flow, single-object short-term tracking, and stereo estimation.

In our project we work on motchallenge.net, an online platform for evalauting methods on a number of well established datasets and challenges. Our website provides a fair and transparent evaluation protocol. The results are shown on a public leaderboard on the website. At this moment, we have +2k active users who already submit +5k methods on challenges reaching from detection, multi-object tracking over tracking and segmentation.

Check out MOTChallenge (here) to download one of the datasets, submit your results or provide data for a new challenge.

The performance of these numerous allows us also to analyze the current trends in tracking methods and gives us an overview of the evolution of the performance of mulit-object trackers. For further read have a look at our IJCV journal paper "MOTChallenge: A Benchmark for Single-camera Multiple Target Tracking".

Pedestrian Trajectory Prediction

Pedestrian Trajectory Prediction is the task of predicting a set of possible future trajectories for a given observation. This ability is indispensable for any autonomous platform navigating through crowded scenes and interacting with humans. In recent years, many research papers have leveraged the power of generative models in order to predict a set of possible future trajectories.

In my research I particulary study multimodality of trajectory predictions. In particualr, I am interested in building models that predict a set of divers but still realistic trajectories.

Goal-GAN

In our ACCV 20 oral paper "Goal-GAN: Multimodal Trajectory Prediction Based on Goal Position Estimation" we build a model that learns a discrete probaility map of future intermediate goal positions. These goal positions are used to pre-condition the decoder to generate the final trajectory.

For more information about Goal-GAN check out the links below:

MG-GAN

In our ICCV 21 paper "MG-GAN: A Multi-Generator Model Preventing Out-of-Distribution Samples in Pedestrian Trajectory Prediction" we propose a multi-generator prediction model that can learns a multimodal trajectory distribution. In the paper, we decompose the task of trajectory prediction into two stages in which we first determine the likelihood of the generator producing realistic trajectories and then predict a trajectory with a selected generator.

For more information about MG-GAN check out the links below:

Student Projects

Next to my main research I constantly supervise students for a guided research project or their theses, especially in the field of pedestrian trajectory prediction. If you are a student from TU Munich and motivated in doing a research project in the field of Pedestrian Trajectory Prediction please contact me via mail.

• Philipp Mondorff: Modeling Social Interactions for Pedestrian Trajectory Prediction on Real and Synthetic Datasets

• David Glavas: Studying Social Interactions in S-GAN