Enhanced driving assistance: automated day and night vehicle detection system utilizing convolutional neural networks

Zaarane, Abdelmoghit; Slimani, Ibtissam; Elhabchi, Mourad; Atouf, Issam · 2024 · Crossref

DOI: 10.11591/ijeecs.v36.i3.pp1532-1542

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical need for reliable vehicle detection in driving assistance systems (DAS) under varying lighting conditions. While many existing systems focus exclusively on daytime detection, the authors note that a significant proportion of road accidents occur at night, yet detection performance often degrades in low-light scenarios. To bridge this gap, the study proposes an enhanced, automated vehicle detection system capable of operating effectively during both day and night using a hybrid approach combining traditional image processing techniques with convolutional neural networks (CNNs). The methodology begins by analyzing the upper third of input images to determine if the scene is captured during the day or night based on average intensity thresholds. For daytime detection, the system employs a two-stage process: hypothesis generation and verification. Hypothesis generation involves converting images to HSV color space, applying histogram equalization and Canny edge detection, and using fast normalized cross-correlation to identify potential vehicle patterns. Verification utilizes a third-level two-dimensional discrete wavelet transform (2D-DWT) to extract features, which are then classified by a CNN. For nighttime detection, the system identifies vehicle lamps through color thresholding in HSV space to isolate bright white (headlamps) and red (taillamps) regions. Connected component analysis extracts these regions, which are then paired based on symmetry using normalized cross-correlation. These candidate pairs are verified by a CNN classifier. The CNN model used is Inception V4, fine-tuned via transfer learning on a custom dataset of 10,000 images sourced from Caltech Cars, AOLP, KITTI, and a nighttime vehicle detection dataset. Experimental results demonstrate high accuracy and real-time performance. The system achieved a precision of 99.2% and a recall of 98.93% for daytime detection, outperforming comparative methods by Ruan et al. and Wei et al. For nighttime detection, the system achieved a precision of 97.57% and a recall of 97.01%, surpassing methods by Kuang et al. and Hemmati et al. The system processed video sequences at an average of 20.72 frames per second for daytime and 22.65 frames per second for nighttime, meeting real-time requirements. The study concludes that this integrated approach significantly improves detection reliability across lighting conditions without requiring additional hardware like lasers, thereby enhancing road safety and reducing accident rates. Future work aims to address performance drops in extremely low-light conditions and expand the dataset for greater generalization.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• dark adaptation mesopic