---
title: Bidirectional Criticality Rating
url: https://www.emergentmind.com/topics/bidirectional-criticality-rating
type: topic
---

# Bidirectional Criticality Rating

Bidirectional criticality rating is a classification methodology designed to enhance the safety evaluation of object detection systems in automated driving by capturing potentially hazardous interactions in both directions between the ego vehicle and other traffic participants. Unlike conventional unidirectional approaches—which assess only from the perspective of the ego vehicle towards its environment—bidirectional rating incorporates mutual dynamics, mitigating the risk of misclassifying critical objects due to asymmetries in vehicle motion or intent. This methodology can be applied to any pairwise criticality metric and operates as a logical-OR extension, yielding substantial improvements in critical object recall and detection timeliness, as empirically validated on safety-critical datasets such as DeepAccident [2512.15181].

## 1. Formal Definition

Let $V = \{ V_1, V_2, \ldots, V_n \}$ represent all agents in a driving scene, with $V_e \equiv V_1$ denoting the ego vehicle. For each pair $(V_i, V_j)$, a pairwise criticality test $M: (V_i, V_j) \rightarrow \{ \mathrm{False}, \mathrm{True} \}$ evaluates whether the inter-vehicle interaction is safety-critical according to a prescribed metric and threshold. The bidirectional criticality label for each non-ego agent $V_i$ ($i > 1$) is given by
$$
C(V_i) = M(V_e \rightarrow V_i) \vee M(V_i \rightarrow V_e)\quad \forall\ i \in \{2, \ldots, n\}
$$
where “$\rightarrow$” denotes the ordered direction of evaluation. $C(V_i)$ thus reflects whether $V_i$ is deemed critical from either the ego's perspective or the reverse.

## 2. Mathematical Formulation

A generic pairwise metric $M(V_i, V_j)$ evaluates a scalar function $f(V_i, V_j)$—for example, time-to-collision (TTC), time-to-arrival (TTA), or Responsibility-Sensitive Safety (RSS) longitudinal distance—against a criticality threshold $T$:
$$
M(V_i \rightarrow V_j) = 
    \begin{cases}
        \mathrm{True} & \text{if}\ f(V_i, V_j) \leq T\quad (\text{or}\ \geq T,\ \text{per metric}) \\
        \mathrm{False} & \text{otherwise}
    \end{cases}
$$
Common instantiations for $f(\cdot, \cdot)$ include:
- **TTC**: $f = \mathrm{TTC} = \frac{d}{|v_i \pm v_j|}$
- **RSS**: $f = d_s$ (required safety distance)
- Other metrics as specified in the literature.

The bidirectional label is then:
$$
C(V_i) = M(V_e \rightarrow V_i; T) \vee M(V_i \rightarrow V_e; T)
$$

## 3. Implementation Procedure

The implementation operates on each frame $t$ as follows:

```python
Input: 
    - Ego state: p_e, v_e, a_e, s_e 
    - Other vehicles: for i > 1: p_i, v_i, a_i, s_i 
    - Metric M and threshold T 
Output: 
    - Criticality labels C_i ∈ {0,1}

For each object i = 2…n:
    1. Compute pairwise states: 
         d_ei ← distance(p_e, p_i)
         Δv_ei ← some function of (v_e, v_i)
         Δa_ei ← some function of (a_e, a_i)
    2. Evaluate M1 ← M(V_e → V_i) using (d_ei, Δv_ei, Δa_ei,…)
    3. Swap roles and compute M2 ← M(V_i → V_e)
    4. Label C_i ← M1 OR M2
Return {C_2, …, C_n}
```

In offline evaluations, ground-truth kinematics are derived directly from annotated datasets; in online or real-time systems, states are estimated from detection and tracking pipelines.

## 4. Extension Beyond Unidirectional Metrics

Unidirectional metrics—those computing only $M(V_e \rightarrow V_i)$—may fail to identify critical interactions in cases such as overtaking, cut-ins, or scenarios where the ego vehicle is not the primary initiator of risk. Bidirectional criticality rating addresses this limitation by performing the same test with reversed roles, capturing:
- Symmetric hazards (e.g., head-on collisions)
- Asymmetric risks where only one perspective detects imminent danger
- Enhanced recall by avoiding false negatives due to relative-motion asymmetry

*This suggests that bidirectional rating generalizes the unidirectional approach, providing a broader and more robust criticality detection framework.*

## 5. Empirical Performance on DeepAccident

Evaluation on the DeepAccident dataset—comprising 159 accident scenarios and 4,995 annotated frames—demonstrates quantitatively superior criticality detection via bidirectional rating. Performance is measured by:
- **SR (Scenario Ratio):** Fraction of scenarios where the colliding object is ever marked critical
- **FR (Frame Ratio):** Fraction of frames the colliding object is marked critical
- **$\tau$ (Time lead):** Time interval between first critical label and collision

| Metric                | Unidir SR | Unidir FR | Bidir SR | Bidir FR | Δ FR   |
|-----------------------|-----------|-----------|----------|----------|--------|
| MTTC @ 4 s            | 0.977     | 0.219     | 0.977    | 0.438    | +100%  |
| RSS (μ=0 m)           | 0.814     | 0.167     | 0.977    | 0.359    | +115%  |
| LSM (Eq. 8)           | 0.953     | 0.583     | 0.953    | 0.662    | +13.5% |
| TTA @ 1.5 s           | 1.000     | 0.976     | 1.000    | 0.996    | +2%    |
| SACRED (all sub-cases)| 1.000     | 0.979     | 1.000    | 0.990    | +1.1%  |

The frame ratio (FR) for weaker metrics such as MTTC and RSS approximately doubles under bidirectional rating, while gains are modest for already-strong metrics (TTA, SACRED). The minimum time-lead $\tau_{min}$ also increases, signifying earlier warning for impending collisions (e.g., for MTTC @4 s from 0.0 s to 0.3 s; for RSS μ=0 from 0.0 s to 0.6 s) [2512.15181].

## 6. Practical Guidance and Recommendations

- Incorporate bidirectional rating with any pairwise criticality metric $M$ to prevent critical scenarios being overlooked due to uni-directional limitations.
- Employ multi-metric aggregation (e.g., logical disjunction of TTC and RSS, or TTA and SACRED) in conjunction with bidirectional evaluation to further elevate SR and FR.
- Calibrate thresholds $T$ individually for each metric and scenario class before applying bidirectional logic for robust adaptation.
- In operational pipelines, estimate vehicle state variables at $\gtrsim$10 Hz when possible, ensuring both $M(V_e \rightarrow V_i)$ and $M(V_i \rightarrow V_e)$ are computed from current tracker estimates.
- Report scenario-level (SR), frame-level (FR), and time-lead ($\tau$) statistics when benchmarking object detection for safety evaluation on accident-prone or near-miss datasets.

By wrapping the existing pairwise metric in a bidirectional logical-OR operation, bidirectional criticality rating yields a systematically more reliable, comprehensive, and accurate method for determining object relevance and hazard in the context of automated driving safety assessment [2512.15181].

Source: https://www.emergentmind.com/topics/bidirectional-criticality-rating