Trajectory Analysis Datasets
The CVRR Trajectory Analysis Dataset provides data for benchmaking trajectory analysis algorithms. The full dataset contains 4 different scenes - a simulated intersection, a real highway, and an 2 indoor omni directional camera sets. The datasets are intended for unsupervised trajectory-based activity analysis which includes classification, prediction, and abnormality detection. The provided trajectories contain only spatial information (velocity must be inferred). True activity labels are provided as well as full track based abnormality label and frame-by-frame unusual event information.
|CROSS: Simulated four way traffic intersection with various through and turn patterns present. Units are pixels. Full labels are available for training and testing for both offline and online analysis.
|I5: Highway trajectories in both direction of I5 outside of UCSD. Trajectories are obtained by a simple visual tracker. Units are pixels. The true cluster labeling considers only the lane. Only trajectory labels are available (no abnormality info).
|OMNI1: Trajectories of humans walking through a lab captured using an omni-directional camera. Units are pixels. Natural trajectories collected over 24 hours on a single Saturday without participant knowledge. Full labels available for training and testing for both offline and online analysis.
|OMNI2: Trajectories of humans walking through a lab captured using an omni-directional camera. Units are pixels. Choreographed trajectories during a 30 minute collection period.
Use the following link to download the dataset: [zip]
Please use the following citation when using the dataset:
B. T. Morris and M. M. Trivedi, "Trajectory Learning for Activity Understanding: Unsupervised, Multilevel, and Long-Term Adaptive Approach," IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 11, pp. 2287-2301, Nov. 2011. [pdf]