One of the biggest challenges of acquiring data for training models in crowd understanding is that ground truth annotations have to be done manually.
But with LCrowdV, the labor intensive effort in annotation and the risk of human mistake is eliminated. Trajectories of every agent, the bounding box of the interested objects, and any other feature one would like to study can be easily generated using our framework.
Each video generated in LCrowdV comes along with 7 labels, which can be varied as parameters.
That includes crowd density, population, lighting conditions, background scene, camera angles, agent personality and noise level.
In other words, the videos produced are in a wide range of variety, with different population density, background environment, individual agent behavior, etc.
We have improved the performance of pedestrian detection using HOG+SVM by 3%, by augmenting the training data using LCrowdV videos.
And we have also combined LCrowdV videos with the training dataset used in Faster R-CNN for pedestrian detection and improved the average precision by 7.3%.
We plan to extend the use of LCrowdV data to different crowd understanding work, including flow estimation, crowd counting, behavior classification, etc.