Phase 1: Related Work (improve)

[wangyashuu]

Suggestion

You could also include some more related work, and try to structure and modularise it.

Stolen Idea:

Two types: Deterministic or Ontological vs Statistical or phenomenological. The former uses physical and differential equations with hard coded rules, such as Social forces model that uses newtonian equations, or RVO; while the latter tries to infer the patterns implicitly and stochastically from within the data, such as by employing neural networks. Due to the complexity of phenomena being modelled, and availability of large data,the latter is preferred.

TODO

make change of our phase1 draft paragraph according to suggestion and idea mentioned above.

[wangyashuu]

Research

approaches

• Physics-based (-> deterministic)
• Pattern-based (-> statistical)
• Planning-based (1 Forward Planning -> deterministic 2. Inverse Planning -> statistical)

conclusion: we can all summerize to deterministic and statistical

[wangyashuu]

Many approaches have been proposed and developed to solve this complex task. In general, they can be categorized as deterministic and statistical. Deterministic methods use hand-crafted functions based on certain observable conditions, such as Newton's laws of motion (which use velocity and acceleration to calculate position) and shortest paths (with assumption that human prefer shortest path to target position), to generate human motion trajectories. A far-reaching example is social forces, a model proposed by Helbing and Molnar~\cite{Helbing95} based on an equation describing the relationship between main effects (including attraction from goal and repulsion from other agents and obstacles) and human motion. Yi~\cite{Yi15} built a model to calculate the optimal path for humans based on the formulated energy map. The statistical ways rely on learning patterns from data through various methods, such as neural networks, Hidden Markov Models, etc. Zhou et al.~\cite{Zhou} build a linear dynamic system, applying Expectation Maximization (EM) algorithm to estimate parameters, to learn motion patterns in crowded scenes. Altché~\cite{Altche17} proposes a method that predicts the trajectory on the highway using Long Short-Term Memory (LSTM). Alahi et al.~\cite{Alahi16} give a sequence model based on LSTM as well as a social pooling that aggregates the human-human interaction in a scene. With the vast amount of data available today, these methods can model complex situations that are difficult for humans to observe, which is valuable information for predicting the behavior of pedestrians. And so, this way is gaining more and more popularity in the research field.

[wangyashuu]

https://github.com/charlietianx/pedestrian-trajectory-pred/pull/16