Dynamic Hazard Field: Methods & Applications

• Dynamic Hazard Field is a collection of time-evolving hazard representations that adapt across domains like crowd simulation, survival analysis, and autonomous driving.
• Its methodologies include explicit Gaussian kernels, ODE-driven hazard functions, QP-MPC risk maps, and PDE-based risk potentials tailored for distinct application needs.
• These frameworks enable practical safety improvements by linking hazard detection to decision-making processes and motion control in complex, dynamic environments.

Searching arXiv for the cited DHF-related papers and related terminology. arxiv_search query: "Dynamic Hazard Field crowd simulation emotional contagion unexpected multi-hazard situations (Xu et al., 2018) survival analysis modelling hazard function via ordinary differential equations (Christen et al., 2023) QP-MPC Dynamic Hazard Field autonomous vehicles (Tian et al., 31 Aug 2025) DRIFT Driving Risk Inference via Field Transmission (Wang et al., 27 May 2026)" Dynamic Hazard Field (DHF) denotes a family of time-evolving hazard representations used in multiple research areas, but it does not refer to a single standardized mathematical object. In crowd simulation, it is a superposed spatial field of multi-hazard danger intensities that is mapped to panic emotion and then to collision-free motion. In survival analysis, it is a hazard function $h(t)$ generated implicitly by an ordinary differential equation system rather than prescribed directly. In autonomous driving, it is a spatial–temporal risk map used for trajectory evaluation and constrained optimization, and, in a more elaborate formulation, a risk potential evolved by an advection–diffusion–reaction partial differential equation with occlusion- and topology-coupled source terms (Xu et al., 2018, Christen et al., 2023, Tian et al., 31 Aug 2025, Wang et al., 27 May 2026).

1. Scope and formal variants

The cited literature uses the term DHF for four closely related but formally distinct constructions. In the crowd-simulation setting, a single-source hazard field is written as a truncated Gaussian kernel,

$H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$

with total field

$H_{\rm total}(x,t)=\sum_{s=1}^n H_s(x,t).$

Here the field is explicitly assembled from active hazard sources, and time dependence enters through the source center $P_s(t)$ and radius $r_s(t)$ when hazards are dynamic (Xu et al., 2018).

In the survival-analysis setting, a DHF is a family of hazard functions defined implicitly by an initial value problem,

$$\frac{dY(t)}{dt}=\psi(Y(t),t;\theta),\qquad Y(0)=Y_0,$$

together with

$H'(t)=h(t),\qquad H(0)=0,$

where $Y(t)=(h(t),q_1(t),\ldots,q_m(t))^\top$. The hazard is therefore a state variable in a dynamical system rather than a fixed functional template such as Weibull or exponential (Christen et al., 2023).

In autonomous-vehicle planning with QP–MPC, the DHF is defined over Frenet coordinates and time as

$H(l,d,t)=H_s(l,d)+H_d(l,d,t),$

where $H_s$ represents static risks and $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$0 moving-obstacle risks. In DRIFT, the analogous object is the instantaneous risk potential

$H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$1

evolved by

$H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$2

with $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$3 in the reported experiments (Tian et al., 31 Aug 2025, Wang et al., 27 May 2026).

A common misconception is that DHF always means a spatial danger map. The survival-analysis formulation shows that the same label is also used for a temporally evolving hazard rate in the statistical sense. This suggests that the unifying feature of DHF is explicit hazard dynamics rather than any single choice of state space or governing equation.

2. Multi-hazard crowd simulation and emotional contagion

In "Crowd Behavior Simulation with Emotional Contagion in Unexpected Multi-hazard Situations" (Xu et al., 2018), hazards are classified along three binary attributes: transient versus persistent, concurrent versus non-concurrent, and static versus dynamic. In the DHF, these distinctions determine whether the active time set $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$4 is a point or an interval, whether Gaussian fields overlap at the same timestep, and whether $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$5 and $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$6 are constant or time-varying. For dynamic hazards, the same kernel is reused with updated parameters,

$H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$7

and the paper also gives a diffusion-based expansion rule,

$H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$8

The field is mapped directly to hazard-induced panic. For an agent $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$9 at position $H_{\rm total}(x,t)=\sum_{s=1}^n H_s(x,t).$0,

$H_{\rm total}(x,t)=\sum_{s=1}^n H_s(x,t).$1

This local perception is then combined with social contagion and natural decay. The contagion dose accumulated over the previous $H_{\rm total}(x,t)=\sum_{s=1}^n H_s(x,t).$2 frames is

$H_{\rm total}(x,t)=\sum_{s=1}^n H_s(x,t).$3

the natural decay term is

$H_{\rm total}(x,t)=\sum_{s=1}^n H_s(x,t).$4

and the total emotion update is

$H_{\rm total}(x,t)=\sum_{s=1}^n H_s(x,t).$5

Motion generation proceeds through the Emotional Reciprocal Velocity Obstacles (ERVO) model. The stress-safety direction is the hazard-weighted sum of away-vectors,

$H_{\rm total}(x,t)=\sum_{s=1}^n H_s(x,t).$6

and the preferred direction is blended with neighbors,

$H_{\rm total}(x,t)=\sum_{s=1}^n H_s(x,t).$7

Within the RVO constraint, the effort-sharing weight becomes emotion-dependent,

$H_{\rm total}(x,t)=\sum_{s=1}^n H_s(x,t).$8

Algorithmically, each timestep updates hazard fields, computes direct panic, accumulates social contagion, applies natural decay, combines the emotion state, builds an escape direction, blends neighbor motion, solves emotion-weighted RVO constraints, and integrates the new position. The significance of this construction is that the field does not only encode environment geometry; it also serves as the upstream driver of a coupled emotion–motion system.

3. ODE-generated hazard dynamics in survival analysis

In "Dynamic survival analysis: modelling the hazard function via ordinary differential equations" (Christen et al., 2023), DHF refers to a parametrically modeled hazard function whose shape is generated dynamically. The central formulation is an autonomous ODE system,

$H_{\rm total}(x,t)=\sum_{s=1}^n H_s(x,t).$9

with $P_s(t)$0. The leading component may be written as

$P_s(t)$1

and the cumulative hazard is part of the same dynamical system through $P_s(t)$2.

The paper emphasizes three advantages. It embeds qualitative “rules of engagement” directly into the hazard model, allows flexible non-monotonic hazard shapes by introducing hidden states, and produces both the instantaneous hazard $P_s(t)$3 and the cumulative hazard

$P_s(t)$4

within one coherent system. The standard survival identities are retained: $P_s(t)$5

Two canonical examples are reported. The logistic-growth hazard,

$P_s(t)$6

admits the closed-form solution

$P_s(t)$7

The Hazard–Response model introduces a hidden response variable $P_s(t)$8: $P_s(t)$9 with $r_s(t)$0. The Lotka–Volterra-style terms $r_s(t)$1 induce non-monotonic, possibly unimodal $r_s(t)$2.

Estimation uses the usual right-censoring likelihood,

$r_s(t)$3

with weakly-informative $r_s(t)$4 priors suggested for positive parameters and posterior computation by Metropolis–Hastings, Metropolis-within-Gibbs, self-tuning t-walk, or HMC, or alternatively by maximum likelihood via optimizers such as nlminb and BFGS. When no closed form exists, the paper recommends numerical initial-value solvers such as LSODA via R/deSolve or SciPy’s solve_ivp, possibly with log-transforms and supplied Jacobians for stability.

The empirical results are concrete. In simulation scenario 1, with sample sizes $r_s(t)$5 and censoring at $r_s(t)$6 or $r_s(t)$7, posterior means and medians converge to true $r_s(t)$8, credible-interval coverage is near nominal $r_s(t)$9, and $$\frac{dY(t)}{dt}=\psi(Y(t),t;\theta),\qquad Y(0)=Y_0,$$0 is the hardest parameter to estimate. In the LeukSurv case study ($$\frac{dY(t)}{dt}=\psi(Y(t),t;\theta),\qquad Y(0)=Y_0,$$1), the logistic-growth hazard yields $$\frac{dY(t)}{dt}=\psi(Y(t),t;\theta),\qquad Y(0)=Y_0,$$2 versus $$\frac{dY(t)}{dt}=\psi(Y(t),t;\theta),\qquad Y(0)=Y_0,$$3. In the Rotterdam breast-cancer case study ($$\frac{dY(t)}{dt}=\psi(Y(t),t;\theta),\qquad Y(0)=Y_0,$$4), the Hazard–Response model gives $$\frac{dY(t)}{dt}=\psi(Y(t),t;\theta),\qquad Y(0)=Y_0,$$5. The paper also states that, compared with a non-parametric B-spline hazard estimator, the DHF posterior predictive hazard is far closer in restricted $$\frac{dY(t)}{dt}=\psi(Y(t),t;\theta),\qquad Y(0)=Y_0,$$6 distance to the truth and is less “wiggly.”

4. Spatial–temporal DHFs for autonomous-vehicle trajectory planning

In "A Risk-aware Spatial-temporal Trajectory Planning Framework for Autonomous Vehicles Using QP-MPC and Dynamic Hazard Fields" (Tian et al., 31 Aug 2025), the DHF is a continuous, spatial–temporal risk map $$\frac{dY(t)}{dt}=\psi(Y(t),t;\theta),\qquad Y(0)=Y_0,$$7 over Frenet coordinates relative to a reference path. It fuses static and dynamic risks: $$\frac{dY(t)}{dt}=\psi(Y(t),t;\theta),\qquad Y(0)=Y_0,$$8 The static field is a Gaussian sum,

$$\frac{dY(t)}{dt}=\psi(Y(t),t;\theta),\qquad Y(0)=Y_0,$$9

where $H'(t)=h(t),\qquad H(0)=0,$0 denotes a static obstacle or boundary point, $H'(t)=h(t),\qquad H(0)=0,$1 its peak hazard level, and $H'(t)=h(t),\qquad H(0)=0,$2 its longitudinal and lateral spreads.

The dynamic field is defined by

$H'(t)=h(t),\qquad H(0)=0,$3

where $H'(t)=h(t),\qquad H(0)=0,$4 is longitudinal separation from the predicted position of moving obstacle $H'(t)=h(t),\qquad H(0)=0,$5, $H'(t)=h(t),\qquad H(0)=0,$6 is obstacle length, $H'(t)=h(t),\qquad H(0)=0,$7 the maximum hazard level, $H'(t)=h(t),\qquad H(0)=0,$8 the relative speed, $H'(t)=h(t),\qquad H(0)=0,$9 a sensitivity gain, and $Y(t)=(h(t),q_1(t),\ldots,q_m(t))^\top$0 a safety coefficient that shifts the logistic curve ahead of the obstacle.

Real-time computation is described as a five-stage process: discretize the Frenet domain and time horizon, precompute $Y(t)=(h(t),q_1(t),\ldots,q_m(t))^\top$1 once per planning cycle, predict obstacle positions using kinematic models, compute $Y(t)=(h(t),q_1(t),\ldots,q_m(t))^\top$2 and add it to $Y(t)=(h(t),q_1(t),\ldots,q_m(t))^\top$3, and then sample the resulting field along each candidate quintic-polynomial trajectory. The hazard exposure term entering the cost is

$Y(t)=(h(t),q_1(t),\ldots,q_m(t))^\top$4

Within receding-horizon QP–MPC, the planner minimizes

$Y(t)=(h(t),q_1(t),\ldots,q_m(t))^\top$5

subject to linearized vehicle dynamics, actuator limits, curvature and jerk limits, and the hard hazard constraint

$Y(t)=(h(t),q_1(t),\ldots,q_m(t))^\top$6

The parameter-selection discussion is explicitly about trade-offs. Static spreads $Y(t)=(h(t),q_1(t),\ldots,q_m(t))^\top$7 are chosen to reflect obstacle size plus a safety margin, dynamic peaks $Y(t)=(h(t),q_1(t),\ldots,q_m(t))^\top$8 and gains $Y(t)=(h(t),q_1(t),\ldots,q_m(t))^\top$9 scale with obstacle speed, and the logistic shift $H(l,d,t)=H_s(l,d)+H_d(l,d,t),$0 is tuned so that $H(l,d,t)=H_s(l,d)+H_d(l,d,t),$1 “lights up” sufficiently in front of each vehicle, with typical $H(l,d,t)=H_s(l,d)+H_d(l,d,t),$2–$H(l,d,t)=H_s(l,d)+H_d(l,d,t),$3. Larger $H(l,d,t)=H_s(l,d)+H_d(l,d,t),$4 yields more conservative, longer paths, slower speed, but smaller $H(l,d,t)=H_s(l,d)+H_d(l,d,t),$5; smaller $H(l,d,t)=H_s(l,d)+H_d(l,d,t),$6 yields faster, more direct trajectories with higher exposure.

The validation examples include lane-changing, overtaking, and intersection crossing. Reported outcomes include zero overlap with the moving-obstacle envelope in the space–time graph for lane-changing, all curvature peaks below the $H(l,d,t)=H_s(l,d)+H_d(l,d,t),$7 comfort limit, average longitudinal jerk dropping from $H(l,d,t)=H_s(l,d)+H_d(l,d,t),$8 under DE to $H(l,d,t)=H_s(l,d)+H_d(l,d,t),$9 under SQP–MPC in overtaking, and collision-free intersection trajectories with a nearly constant speed profile. The paper states that the DHF-augmented framework outperforms DE, PSO, Interior-Point, and Active-Set in efficiency, stability, and comfort.

5. PDE-based DHFs in DRIFT

DRIFT formulates the DHF as a spatiotemporal risk field governed by transmission dynamics rather than frame-wise superposition alone (Wang et al., 27 May 2026). In the zero-inertia setting used in the experiments, the governing PDE is

$H_s$0

The initial condition is $H_s$1 for all $H_s$2, the ego-sensor boundary uses zero-flux Neumann conditions $H_s$3, and a downstream sponge layer adds a large $H_s$4 to absorb advected risk.

Observed vehicles contribute anisotropic Gaussian kernels,

$H_s$5

with rotated covariance

$H_s$6

Typical values are $H_s$7 and $H_s$8. The source weight may be modulated by speed and acceleration as

$H_s$9

with $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$00, $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$01, $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$02, and $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$03. Braking kernels are often shifted forward by $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$04, with $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$05.

Occlusion-aware latent hazards are injected through a shadow mask $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$06 and occluded region $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$07: $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$08 As the mask expands, $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$09 inside $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$10 is reduced so that latent risk accumulates; as the mask retreats, $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$11 is boosted to clear now-visible cells. Topology-coupled merge conflict pressure is modeled by

$H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$12

with typical $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$13 and $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$14. The total source term is $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$15.

Numerically, DRIFT uses a fixed BEV grid of size $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$16 at $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$17, with $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$18 and $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$19. Diffusion is discretized by central differences, advection by upwind differencing, and time by explicit Euler: $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$20 The stability condition is

$H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$21

DRIFT also introduces four field-centric evaluation metrics: Lane-Change Risk Differential (LCRD), Temporal Anticipation Index (TAI), Occlusion Sensitivity Index (OSI), and Occlusion Response Latency (ORL). On exiD, rounD, and inD, the reported comparison shows DRIFT achieving $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$22, $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$23 (equivalently $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$24), $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$25, and $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$26, compared with RiskNet at $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$27, $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$28 ($H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$29), $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$30, and $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$31, and EDRF at $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$32, $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$33 ($H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$34), $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$35, and $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$36. The near-collision increase under occlusion is reported as $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$37 for DRIFT, versus $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$38 for RiskNet and $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$39 for EDRF.

6. Interpretation, limitations, and recurring points of confusion

Several recurring points of confusion can be resolved directly from the literature. First, DHF is not a single canonical equation. In the crowd-simulation formulation it is a superposition of truncated Gaussian hazard kernels; in survival analysis it is an ODE-defined hazard trajectory; in QP–MPC it is a Frenet-coordinate risk map combining Gaussian static terms and logistic dynamic terms; in DRIFT it is a PDE-evolved risk potential with advection, diffusion, reaction, occlusion sources, and merge pressure (Xu et al., 2018, Christen et al., 2023, Tian et al., 31 Aug 2025, Wang et al., 27 May 2026).

Second, “dynamic” does not always mean that hazards physically move through space. In the crowd paper, dynamic hazards do mean time-varying $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$40 or $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$41. In the autonomous-driving papers, dynamicity includes predicted moving obstacles and temporal risk propagation. In the survival-analysis paper, by contrast, the word refers to the temporal evolution of the hazard function itself under mechanistic laws.

Third, DHFs differ in how they connect perception to decision-making. In crowd simulation, the field is mapped to panic emotion and then to ERVO-based motion. In QP–MPC, it enters the planner as both an integral cost and a hard constraint. In DRIFT, it functions as a field-native safety representation evaluated with metrics designed to complement TTC and PET. In survival analysis, the field is not used for control at all; it is an inferential model for event-time distributions.

These differences also delimit the main limitations stated in the papers. The QP–MPC formulation explicitly depends on parameter tuning for $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$42, $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$43, and $H_s(x,t)=\Gamma_s(x,t)= \begin{cases} \displaystyle \frac{1}{\sqrt{2\pi}\,r_s}\, \exp\Bigl(-\frac{\|x-D_s(x)\|^2}{2\,r_s^2}\Bigr) & \text{if } \|x-D_s(x)\|<r_s\;\wedge\;t\in U_s,\[1em] 0 & \text{otherwise,} \end{cases}$44. DRIFT is motivated by the observation that hand-crafted static risk field models struggle with occlusion and topology-driven propagation. The survival-analysis paper notes that some DHF systems require numerical solvers and that bias in the Hazard–Response simulation study can arise from grid coarseness, though this is remedied by finer grids. A plausible implication is that DHF is best understood as a modeling pattern for explicit hazard dynamics, whose practical value depends on how well the chosen governing law matches the causal structure of the application domain.