Activity Forecasting

Modeling High-Dimensional Humans for Activity Anticipation using Gaussian Process Latent CRFs

“…In another previous work of anticipating human activities [20,19], human motions are implicitly modeled through object trajectories. Some other works [3,18,44,23] predict possible human navigation trajectories in 2D from visual data. In all the aforementioned works, human motions are under-represented and would fail when a more elaborate prediction on human movements is required.…”

“…Therefore, some works assume a few static human poses are representative enough [6,8,13,12] or simplify a human configuration to a 2D point for navigation task [3,18,44,23] or to a 3D trajectory of one hand while keeping the rest body static neglecting kinematic constraints [20,19]. In these works, human motions are underrepresented and would fail when a more elaborate human motion prediction is required.…”

Tracking and intercepting pedestrians: a robotic approach to surveillance of critical infrastructures

“…However, a relative estimation could be calculated based on the difference between maximum and minimum MHD. The work of Kitani et al [83] has a relative reduction of 75% and the results presented produce an improvement of 86%. Although this is not conclusive, it shows that the proposed technique produces a useful prediction, even when working in large scenarios.…”

“…Finally, in order to compare the proposed prediction with previous works, a formal evaluation of the results presented here can be obtained by computing the Modified Hausdorff Distance (MHD) as it is done by Kitani et al [83]. This measurement gives standardized a-posteriori information about the similarity of the predicted routes and the trajectory followed by the pedestrian.…”

“…Then, it creates a cost-map based on previously observed trajectories, and uses it to plan a future trajectory of the newly observed pedestrian. This work was later extended by adding vision based physical scene features and noisy tracker observations [83]. Those works are similar to the one proposed in this paper in the sense that they also use a planning step to predict the future position of the pedestrian, however, their long-term prediction is not very clear, because it only directs its future position and therefore it is only valid in relatively short distances, also, time or velocity issues are not accounted for, and as with all other prediction techniques they require a large amount of observations.…”