PVM (Poor Visibility Module) Real-Time Glare Detection for Sun-Glare, Night-time Glare

Oct 9, 2025

Current State

Industry and Issue

​

2

Current State

Detection systems of all types have increased accuracy in all weather conditions

The gap between ‘the floor and the ceiling’ for detection is narrower than ever

Assessing different detection methods involves objective and subjective criteria

Assessing the benefits of the ‘full package’ is more important than ever

3

Problem

Detection reliability for video can be compromised by challenging visual conditions, specifically sun-glare and night-time glare.

​

Current glare detection methods rely on intensity-based filtering and other classical computer vision techniques, which lack the necessary robustness.

4

Problem Statement

• Our Goal is to enhance the system’s ability to identify these conditions accurately, and generate real-time alerts, by leveraging advanced deep learning approaches.

5

Goal

To enhance the system’s ability to identify these conditions accurately, and generate real-time alerts by leveraging advanced deep learning approaches

• Be more accurate at all times, and reduce window of ‘failing high’

6

Solution

​

​

7

Deep Learning based Poor Visibility Detection

• Employ a fully convolutional neural network approach, trained on data from hundreds of intersections (~30,000 real samples) to create an efficient model designed for robust glare detection.

8

Deep Learning based Poor Visibility Detection

• Ensure no false positives and thus no false alerts by:
• Using temporal hysteresis on detected glare and a rolling time window to ensure only persistent glare triggers the alert
• Including cascaded filtering layers on the detected glare
• Proposed solution is perspective-invariant and highly generalizable

9

Outcome

• Achieved: Developed a robust deep learning–based glare detection model that performs reliably in both day and night, under extreme as well as moderate conditions.
• Improved: Overcame the challenge of high false positives during daytime, delivering much stronger accuracy across diverse scenarios.
• Benefit: Provides consistent and accurate glare detection across all lighting conditions, significantly improving system reliability and safety. �

​

10

Next Steps

• Complete Integration: Finalize pipeline integration to enable seamless deployment.
• Thorough Validation: Run soak tests and evaluate corpus results for regression to thoroughly validate the approach on existing data.
• Beta Rollout:
1. Deploy the new model at 3 intersections, at first, as a controlled trial and expand to 10 intersections (Targeting initial Beta by October, 2025);
2. Monitor performance for 1 month with automatic fallback to the previous model if needed; upon success, expand to 5 intersections and progressively scale further.
• General availability: Release the solution as a feature, making glare detection a standard part of our ecosystem.�

​

11

Daytime Extreme to Moderate Glare (PIMA)

​

12

Daytime Extreme to Moderate Glare (PIMA)

13

No Glare (PIMA)

​

14

No Glare

​

15

Partial Glare (PIMA)

​

16

Night-time Glare

​

17

Night-time Glare

​

​

18

Night-time Glare Wet Road

​

​

19

QnA

​

​

20