Robust Lane Line Detection for Intelligent Vehicles under Complex Illumination Conditions Based on Image Processing

With the rapid development of intelligent driving technology, the autonomous navigation capability of intelligent vehicles in dynamic environments has become a research hotspot [1-3]. Visual navigation, as the core method of the perception system of intelligent vehicles [4, 5], directly affects the vehicle’s path planning and decision control. However, in actual driving scenarios, illumination conditions are complex and variable, such as strong direct light [6], shadow occlusion [7], backlight environments [8], etc., which often cause problems such as reduced image contrast [9], increased noise [10], and feature blurring [11], seriously affecting the accuracy and robustness of traditional recognition algorithms. How to achieve high-precision lane line recognition under complex illumination conditions has become a key technical bottleneck restricting the widespread application of intelligent vehicles in unstructured environments.

This study has important theoretical significance and engineering value for improving the environmental adaptability of intelligent vehicles under complex illumination scenarios. From a theoretical perspective, by exploring the mapping relationship between illumination changes and image features and constructing a robust image processing model, it can enrich the theoretical system of visual perception for intelligent vehicles; from the engineering application perspective, accurate lane line recognition is the foundation for intelligent vehicles to achieve autonomous obstacle avoidance and lane line tracking. The research results can effectively improve driving safety and reliability of vehicles under all-weather conditions and provide technical support for the industrialization of autonomous driving technology.

Existing lane line recognition methods for intelligent vehicles have obvious limitations under complex illumination conditions. Traditional threshold segmentation-based methods [12, 13] require high illumination uniformity and are prone to mis-segmentation in shadow areas; edge detection-based algorithms [13, 14] are sensitive to contours but easily produce false edges under strong light interference. In recent years, deep learning-based methods [15-17] perform well under conventional illumination but suffer from degraded feature extraction capability due to image information degradation in low-light or strong-light scenarios. Moreover, these models have large numbers of parameters and high computational complexity, making it difficult to meet the real-time requirements of intelligent vehicles. Additionally, some studies use HE for image enhancement [18-20], but this method easily amplifies noise and has insufficient adaptability to dynamic illumination changes.

This paper focuses on the lane line recognition problem of intelligent vehicles under complex illumination conditions and mainly conducts two aspects of research: on the one hand, proposes an image enhancement method based on optimized MSR, improving the illumination component estimation and reflectance component recovery process to enhance the contrast and detail information of lane line images under complex illumination; on the other hand, constructs a lane line recognition algorithm based on SSA-optimized spectral clustering, utilizing the global optimization capability of the SSA to optimize the similarity matrix construction and cluster center initialization process of spectral clustering, improving the distinguishability between lane lines and background. The innovation of the research results lies in combining image enhancement and intelligent optimization algorithms to form an integrated “enhancement-recognition” solution, which can effectively suppress the influence of illumination changes and improve the robustness and real-time performance of the recognition algorithm, providing a new technical path for reliable navigation of intelligent vehicles in complex illumination environments.

3.1 Spectral clustering algorithm

Under complex lighting conditions, intelligent vehicle lane line images often present multimodal feature distributions of lane line and background due to shadow occlusion, strong light reflection, and other factors, such as asphalt pavement, white lane lines, tree shadows, and road reflections, forming different categories. Traditional two-class spectral clustering criteria, due to unbalanced segmentation problems, are difficult to accurately separate lane lines from multi-class backgrounds in complex scenes. In addition, image noise and contrast fluctuations caused by complex illumination lead to non-uniform distribution of lane line pixel features. Using two-class division easily causes misclustering of lane line edges with similar backgrounds. Therefore, this paper adopts multi-way partition criteria to meet the multi-class segmentation requirements of "lane line - strong light area - shadow area - background" in intelligent vehicle lane line recognition. By dividing image pixels into multiple categories, it adapts to the multimodal distribution of lane line features under complex illumination, avoiding recognition errors caused by information simplification in two-class division. Figure 3 shows the arc error schematic of intelligent vehicle lane line recognition.

3.png

Figure 3. Arc error of intelligent vehicle lane line recognition

The basic principle of the multi-way partition criterion is to construct a multi-way normalized cut model based on spectral graph theory, regarding the intelligent vehicle lane line image as an undirected graph, where pixel points are nodes and similarity between nodes is the edge weight. By solving the generalized eigenvalue problem, clustering eigenvectors are obtained. Specifically, the multi-way normalized cut criterion balances the number of nodes and the cut edge weights of each subgraph, avoiding the "small graph overcut" problem caused by the traditional minimum cut. In complex lighting scenes, this mechanism can ensure that lane line pixels in shadows, strong light, and different background categories are reasonably divided, preventing misclassification of lane line pixels into the background due to lighting differences. Assume the two subgraphs partitioned from the mother graph are represented by X and Y, with vertices of subgraphs X and Y represented by i and n, the weight between subgraphs denoted by q, and the similarity between the two subgraphs represented by CUT(X, Y), then there is the expression:

$\operatorname{CUT}(X, Y)=\sum_{i \in X, n \in Y} q(i, n)$                          (10)

Assuming subgraph u is represented by X_u, and the remaining subgraphs except u are represented by $\bar{X}_u$, the sum of the weights of edges connecting all data samples within subgraph X_u is represented by AS(X_u,X_u), and the maximum minimum cut is represented by LVCUT, then there is the expression:

$L V C U T=\sum_{i=1}^j \frac{\operatorname{CUT}\left(X_u, \bar{X}_u\right)}{\operatorname{AS}\left(X_u, X_u\right)}$                          (11)

3.2 Intelligent vehicle lane line recognition model

Under complex lighting conditions, the pixel features of intelligent vehicle lane line images exhibit highly non-uniform distributions due to interference such as strong light and shadows. Traditional spectral clustering algorithms cannot adaptively determine the number of clusters K and the key parameter v, which easily leads to mis-clustering between lane lines and the background. For example, when the vehicle drives from a tunnel into a sunlit road section, the sudden change of illumination causes drastic variations in the brightness and chromatic features of lane line pixels. If spectral clustering with fixed parameters is used, it may mistakenly classify lane lines under strong light as reflective background areas. Therefore, this paper proposes a SSA-optimized spectral clustering model. Its core idea is to use a heuristic algorithm to compensate for the parameter dependence deficiency of traditional spectral clustering, utilizing the global optimization capability of the SSA to dynamically adapt to the multi-modal distribution of lane line features under complex lighting, achieving adaptive optimization of clustering parameters and avoiding disconnection between manual parameter tuning and scenario changes. Figure 4 shows the schematic diagram of the intelligent vehicle lane line model constructed in this paper.

4.png

Figure 4. Intelligent vehicle lane line model schematic

The implementation process of this method closely revolves around the recognition needs of lane lines under complex lighting: first, randomly initialize the initial clustering centers for pixel features enhanced by optimized MSR, constructing initial partitions including lane lines, strong light areas, shadow areas, background, and other categories; then introduce the SSA, using the silhouette coefficient as the fitness function, simulating the foraging and vigilance behaviors of the sparrow population to iteratively optimize the spectral clustering similarity matrix construction parameter v and the number of clusters K. In scenarios with variable lighting, SSA can adjust parameters in real time. For example, when large shadow areas appear in the image, it automatically increases K to subdivide lane lines and dark backgrounds within the shadows, while optimizing v to enhance the connection weights of similar pixels; finally, through multi-way spectral graph partitioning under the optimal parameters, the lane lines are accurately separated from the complex background. Suppose the average dissimilarity between an intelligent vehicle lane line u and other lane lines in the same cluster is represented by R(O_u), and the minimum dissimilarity between u and lane lines in other clusters is represented by MIN(e_u), then the silhouette coefficient calculation formula is given by:

$S C_n=\frac{M I N\left(e_n\right)-R\left(O_n\right)}{M I N\left\{R\left(O_n\right)-M I N\left(e_n\right)\right\}}$                          (12)

Suppose the trace of a matrix is denoted by tr, the between-class covariance matrix is S_v, the within-cluster covariance matrix is P_v, the number of samples is L, and the number of classes is V, then the CH index calculation formula is given by:

$C H=\frac{t r\left(S_v\right)}{t \cdot\left(P_v\right)} \frac{L-V}{V-1}$                         (13)

The steps of intelligent vehicle lane line recognition based on SSA-optimized spectral clustering are as follows, with the algorithm flowchart shown in Figure 5.

5.png

Figure 5. Intelligent vehicle lane line recognition algorithm flowchart

Step 1: Data Resampling and Preprocessing

Firstly, the original lane line images are resampled. Using cubic spline interpolation and other time synchronization strategies, images under different lighting conditions are unified to the same sampling density to weaken the feature fluctuations caused by dynamic lighting changes. The resampled data undergoes preprocessing and normalization to eliminate the direct influence of light intensity differences on pixel values. For example, in backlight scenarios, normalization can map the overexposed sky region and darker lane line pixels to a unified value range, avoiding feature shifts caused by large brightness spans. Based on optimized MSR enhanced image features, a fused similarity distance matrix E is constructed, which comprehensively considers the YUV luminance component and RGB color features. In complex lighting scenes with mixed shadows, it effectively measures the similarity between lane line pixels and the background.

Step 2: SSA Parameter Initialization

For the dynamic changes of lane line features under complex lighting, initialize the key parameters of the SSA. The population size v should balance computational efficiency and optimization capability; in scenarios of drastic light changes, a larger population size improves coverage of multi-modal features. The iteration number l must adapt to the frequency of lighting changes to ensure real-time parameter updates during high-speed driving. The alert threshold ts relates to sensitivity to sudden illumination changes; when the image brightness gradient exceeds ts, the vigilant sparrows trigger parameter adjustment to avoid clustering parameter failure due to abrupt light changes. The search dimension DIM corresponds to key parameters of spectral clustering, and its dimensionality should match the multi-class segmentation needs of lane line recognition under complex lighting.

Step 3: Fitness Function Definition and Optimization Preparation

Use the silhouette coefficient SC as the fitness function, whose core role is to quantify clustering quality under complex lighting conditions, solving the problem of manual tuning dependence in traditional spectral clustering parameters. In uneven lighting lane scenarios, SC effectively evaluates the tightness of lane line pixels within their clusters and the separation from other clusters. For lane lines heavily affected by reflections under strong light, SC can identify blurred boundaries with reflective background regions, improving distinguishability by parameter optimization. When setting the optimization parameter ranges, the feature fluctuation range caused by lighting changes should be considered: the upper limit of the number of clusters K must accommodate the multi-modal distribution under extreme lighting; the adjustment range of parameter v must cover pixel similarity calculation demands under different lighting intensities. SSA's global optimization ability computes the fitness value corresponding to each sparrow individual's parameter combination, providing quantitative support for dynamically adapting to lighting changes.

Step 4: Sparrow Population Position Update

During iterative optimization under complex lighting, the top V sparrows with better fitness values act as discoverers responsible for exploring new search areas in the parameter space, coping with sudden lighting changes. The discoverers' position update formula incorporates sensitivity factors to lighting changes; when image brightness variance exceeds the threshold, the exploration step size increases to quickly search for new parameter combinations. The remaining sparrows act as followers, following the optimal positions of the discoverers to fine-tune parameters in stable lighting regions. Simultaneously, a subset of sparrows is randomly selected as scouters. Their position update formula couples with the lightning sudden change detection mechanism; when local strong light spots or large shadow areas appear in the image, the scouters trigger parameter reset mechanisms to prevent clustering bias caused by local lighting anomalies, ensuring accurate recognition of lane lines under sudden lighting changes. Suppose the discoverers' and followers' individual positions are represented by N^s+1 and N^s+1_c, the current iteration number by u, a normally distributed random number by W, the maximum iteration number by l, xs belongs to [0.5,1], the matrix R of ones with size 1×f, the population size v, the best discoverer position N_o, the current global worst position N_BAD, and the matrix X with elements 1 or -1, then the position update formulas for discoverers and scouters are given by:

$N^{s+1}=\left\{\begin{array}{l}N^s * \exp \left(\frac{-u}{\partial^* l}\right), t s<x s \\ N^s+W^* R, t s>x s\end{array}\right.$                         (14)

$\begin{aligned} & N_c^{s+1}=\left\{\begin{array}{l}W^* \exp \left(\frac{N_{B A D}^s-N^s}{u^2}\right), u<v / 2 \\ N_o^{s+1}+\left|N^s-N_o^{s+1}\right| * X^T\left(X X^T\right) * R, u>v / 2\end{array}\right.\end{aligned}$                            (15)

Step 5: Population Best Position Update

In each iteration, by comparing the fitness values of all sparrows in the current population, update the global best position c_a and the corresponding best fitness c_d. This mechanism is crucial in complex lighting environments. For example, when vehicles continuously pass through multiple bridge tunnels with alternating light and shadow, the algorithm can accumulate the best parameters from historical iterations, forming a memory ability for periodic lighting changes, thereby improving the optimization speed of parameters for subsequent similar scenarios. The best position c_a corresponds to the parameter combination that can dynamically adapt to feature distribution changes caused by lighting: in low-light scenes, increase the K value to subdivide the lane lines in shadows from the dark background; in strong light scenes, adjust the v value to weaken the similarity weight of reflective pixels, avoiding confusion between lane lines and background. By continuously updating the best position, the algorithm can maintain robust recognition of lane lines during driving with constantly changing lighting.

Step 6: Iteration Termination Condition Judgment

The setting of the iteration termination condition needs to balance computational real-time performance and parameter optimization accuracy to meet the requirements of high-speed intelligent vehicle scenarios. When the number of iterations reaches the preset value l or the change amplitude of the best fitness c_d is less than the threshold, the optimization process terminates. In stable lighting scenes, the algorithm can converge early to reduce computational cost; in scenes with drastic lighting changes, the iteration number is automatically extended to ensure sufficient parameter optimization.

Step 7: Spectral Clustering Matrix Construction under Optimal Parameters

The best parameters obtained by SSA optimization are used in the spectral clustering algorithm to construct the adjacency matrix for the lane line image enhanced by MSR. The K-nearest neighbor method is used to calculate the similarity between each pixel and its nearest K points. In complex lighting scenes, this method can adaptively adjust the adjacency range: in strong light reflection areas, reduce the K value to shrink the adjacency range and avoid erroneous connections between reflective pixels and lane lines; in shadow blur areas, increase the K value to enhance the connection of weak feature pixels, ensuring the continuity of lane line edges. Based on the adjacency matrix, the degree matrix F and Laplacian matrix L are constructed. The construction of the Laplacian matrix incorporates lighting adjustment factors, applying similarity attenuation weights to pixels in strong light areas and similarity enhancement weights to pixels in shadow areas, thereby suppressing the interference of lighting changes on clustering at the matrix level. Assume the intelligent vehicle lane line samples are represented by a_u, a_k, the specific formulas are as follows:

$Q_{u k}=\left\{\begin{array}{l}\exp \left(-\frac{\left\|a_u-a_k\right\|^2}{2 \delta^2}\right),(1) \\ 0,(2)\end{array}\right.$                          (16)

Assuming the sum of each row of the adjacency matrix Q is f_u. If a_u∈KNN(a_k) and a_k∈KNN(a_u), then formula (16) applies, otherwise formula (17) applies.

$F=\operatorname{DIAG}\left(f_1, f_2, f_3, \ldots, f_v\right)$                          (17)

$f_u=\sum_{k=1}^v Q_{u k}$                         (18)

Step 8: Feature Vector Extraction and Normalization

According to the optimal number of clusters K, extract the first K largest eigenvalue corresponding eigenvectors of the normalized Laplacian matrix L to construct a v×K dimensional feature matrix B. Under complex lighting conditions, this step reduces dimensionality to eliminate high-dimensional interference caused by lighting noise. For example, map lane line pixels under strong light from the RGB-YUV mixed feature space to low-dimensional feature vectors, highlighting their essential differences from the background. The normalization of feature vectors further balances the scale differences caused by lighting changes: in low-light scenes, enhance the weight of brightness features; in strong light scenes, suppress the abnormal contribution of over-bright pixels, making lane line features comparable under different lighting conditions.

Step 9: k-means Clustering and Lane Line Segmentation

Use the k-means algorithm to cluster the feature matrix B, dividing pixels into K clusters to separate lane lines from the background under complex lighting. During clustering, the initial cluster centers are initialized by the SSA optimization results, avoiding clustering bias caused by traditional random initialization in sudden lighting change scenarios. For example, at the strong light and shadow junction of a tunnel exit, the optimized initial centers can accurately capture the bimodal features of the lane line. The distance metric in clustering incorporates lighting adaptive weights: for pixels in strong light areas, use a combination metric of Euclidean distance and brightness penalty; for pixels in shadow areas, enhance the metric with color similarity, so that lane line pixels under different lighting are grouped into the same class, while background pixels with similar lighting are grouped into other classes. Finally, each row in Y corresponds to the cluster membership of the pixel, directly reflecting whether it belongs to the lane line.

Step 10: Lane Line Center Extraction and Recognition

Based on the clustering results, use the characteristic that lane lines have density maxima even under complex lighting to extract the central lane line as the recognition basis. Specifically, calculate the average sum of feature fusion Euclidean distances of pixels in each cluster, and determine the cluster corresponding to the minimum value as the central cluster of the lane line. When strong light reflection causes blurred edges of the lane line, density analysis can exclude reflective noise interference to locate the true center; when shadows cause lane line breaks, density connectivity can restore continuous central lane lines. The formula for extracting the central lane line incorporates lighting adjustment factors, applying attenuation coefficients to distance calculations for pixels in strong light areas and enhancement coefficients for pixels in shadow areas, so that in uneven lighting scenes, the center extraction results more closely match the actual lane line position, providing a reliable visual basis for intelligent vehicle path planning. Assume the lane line in cluster k is represented by H_k, and the remaining lane lines in the cluster by $\bar{H}_k$, the specific formula for extracting the central lane line is as follows:

$C_{M I N}=M I N\left\{\sum D I S\left(H_k, \bar{H}_k\right) / v\right\}, k \in(1, v)$                           (19)