Giving Cars Eyes: Using Computer Vision to Prevent Road Traffic Accidents in Nigeria

Chidiogo NwabuikeAAAI 2025 · DOI

Abstract

Nigeria faces a critical road safety crisis, ranking second globally in traffic accident rates. This paper proposes a proactive technological intervention by integrating Dynamic Vision Sensors (DVS) with deep learning architectures to predict steering commands in real-time. By utilizing event-based cameras that capture pixel-level changes rather than static frames, the proposed system offers superior performance in high-speed and low-light conditions typical of Nigerian urban environments. Preliminary results using the DDD17 dataset demonstrate that event-based processing significantly improves steering angle regression accuracy compared to traditional frame-based methods, providing a viable path toward SAE Level 2 automation for existing vehicle fleets.

Introduction

With a population exceeding 200 million and over 11.8 million vehicles, Nigeria’s road infrastructure is under immense pressure. The human and economic cost of traffic fatalities is staggering, yet most vehicles in the region remain at SAE Level 0 or 1, offering little to no automated safety assistance. Furthermore, existing autonomous driving models are predominantly trained on Western datasets, which fail to account for the unique "road irregularities" and chaotic traffic patterns of cities like Lagos. This research seeks to bridge this gap by developing a computer vision framework tailored to the Nigerian context, focusing on real-time hazard identification and proactive steering assistance.

Methodology: Beyond Frame-Based Vision

Traditional computer vision relies on frame-based cameras that capture full images at fixed intervals. This approach often results in high data redundancy and poor performance in challenging lighting or high-speed scenarios.

Event-Based Processing

The proposed system leverages Dynamic Vision Sensors (DVS), or event cameras. Unlike standard sensors, DVS pixels operate asynchronously, triggering only when they detect a change in brightness.

Visualization of an event-based frame where pixel changes are captured asynchronously, highlighting edges and motion while ignoring redundant background data.

This results in microsecond temporal resolution, high dynamic range for low-light performance, and significant computational efficiency by processing only the sparse stream of changes rather than entire image grids.

System Architecture

The framework is partitioned into two functional blocks:

  1. Perception System: Handles localization, road mapping, and the tracking of both static and moving obstacles.
  2. Decision-Making System: Responsible for high-level route planning and path selection based on perception data.

A deep neural network is employed to map the sparse "event frames" directly to steering wheel angles. Performance is evaluated using Absolute Trajectory Error (ATE) and Relative Pose Error (RPE).

Datasets and Training

The model is trained and validated using the DDD17 (DAVIS Driving Dataset 2017), which provides synchronized event-based visual input and ground truth steering angles from real-world driving. While foundational datasets like ImageNet and COCO inform the underlying deep learning techniques, the author emphasizes the need for a localized "Nigerian datasheet" to capture regional environmental factors.

Results and Discussion

Qualitative analysis reveals that the event-based model significantly outperforms traditional frame-based approaches, particularly in complex lighting conditions.

Comparison of steering prediction performance between standard frame-based cameras (left) and event-based sensors (right) in both day and night conditions.

As shown in the comparison above, standard frame-based models (left column) struggle to accurately regress steering angles, showing significant deviation from ground truth in both day and night scenes. In contrast, the event-based model (right column) achieves high precision, with predicted angles closely mirroring the ground truth (e.g., predicting -10.5 degrees against a -10.2 degree ground truth in low-light conditions). This efficiency through sparsity allows the system to ignore redundant background data and focus on critical movement.

Conclusion and Future Work

Transitioning the Nigerian vehicle fleet from Level 0/1 to Level 2 automation—specifically through proactive steering assistance—could potentially reduce road accidents by one-third. However, several challenges remain. Sensors must be equipped with robust casings to withstand the physical shocks and climatic conditions of the region. Furthermore, future work will focus on experimental data collection in suburban Lagos to refine the model for local road conditions. Increasing public and academic engagement with AI in Nigeria is crucial for the long-term sustainability of these safety technologies.

Summary updated

BibTeX
@inproceedings{nwabuike2025giving,
  author = {Chidiogo Nwabuike},
  title = {{Giving Cars Eyes: Using Computer Vision to Prevent Road Traffic Accidents in Nigeria}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35336},
}
Figure