Material Identification of Scanned Objects Based on the Classification of the Laser Reflection Intensity Profile

TL;DR

This paper introduces a new way to identify materials like glass using laser scanner data in robot navigation.

Contribution

The novel approach uses laser reflection intensity profiles and gradients for material classification without needing multiple scans or sensors.

Findings

The method successfully classifies materials including transparent glass using single-frame LiDAR data.

Using intensity gradients improves classification accuracy in overlapping material regions.

Experimental results show reliable performance in indoor environments.

Abstract

This paper presents a method for material classification of objects detected by a laser scanner (LiDAR) used in autonomous mobile robot navigation. The proposed approach operates on a single-frame LiDAR scan composed of single-beam echoes and addresses materials with different reflective properties, including transparent glass surfaces. Material classification is performed by comparing measured reflection intensity profiles, defined as functions of distance and beam incidence angle, with reference profiles constructed for selected material classes. In addition to normalized reflection intensity, the gradient of the intensity profile is used to support discrimination in regions where material-dependent characteristics overlap. Experimental results obtained in indoor environments containing glass surfaces demonstrate that the proposed method enables reliable material type classification without multi-scan data accumulation or multi-sensor fusion.

Figures

Full text

1. Introduction

Autonomous mobile robots are machines capable of operating independently in a given environment without causing collisions or accidents and without direct human intervention. This autonomy is achieved through the use of simultaneous localization and mapping (SLAM) algorithms, which provide information about the robot’s pose with respect to static and dynamic obstacles present in the environment. In recent years, mobile robots have been increasingly deployed in indoor applications, such as transport, cleaning, and inspection tasks, as well as in support operations carried out in hazardous situations (e.g., fire or bomb incidents).

Autonomous navigation of mobile robots has been extensively studied in the context of indoor and structured environments, where reliable perception of surrounding obstacles is critical for safety and task execution [1,2]. In such systems, exteroceptive sensors such as laser scanners and cameras play a central role in environment perception and mapping.

In such environments, including shopping malls and office buildings, robots encounter objects made of materials with diverse optical properties. These include highly reflective materials, such as polished metal surfaces or mirrors, as well as transparent materials for visible light, such as glass windows and glass doors. Objects of this type are particularly difficult to detect using light-based sensors commonly employed in SLAM systems for autonomous navigation.

Laser scanners and RGB or RGB-D cameras are frequently used for this purpose; however, these sensors operate in different spectral ranges. While RGB and RGB-D cameras rely on visible or near-visible light, most commercially available LiDAR sensors operate in the near-infrared spectrum (typically around 850–905 nm), which affects the interaction of laser beams with various materials, particularly transparent and reflective surfaces [3]. For the material types considered in this work (glass, metal, and matte surfaces), these wavelength-dependent effects primarily influence reflection intensity rather than geometric measurement accuracy.

In the case of active range sensors such as LiDAR, transparent glass surfaces are reliably detected only when the laser beam strikes the surface at an incidence angle close to zero. At larger incidence angles, the reflected signal intensity may become too weak to be registered by the sensor, resulting in only a fragment of the scanned object being visible to the navigation system. When measurements obtained from different robot poses are matched during map construction, single-beam echoes from glass surfaces may therefore be interpreted as dynamic disturbances, because different parts of the object are visible from different measurement locations. This phenomenon makes it difficult to create a correct and consistent map of the environment, which in turn degrades localization accuracy and complicates reliable motion planning.

Vision-based sensing modalities exhibit analogous limitations. Standard RGB cameras operate in visible light, which is largely transmitted through transparent glass rather than reflected, rendering such objects effectively invisible in camera images. RGB-D cameras, which provide depth information in addition to color data, also experience difficulties when interacting with transparent materials. In these systems, the quality of depth estimation depends strongly on viewing geometry and illumination conditions, and performance may degrade significantly in the presence of natural light. Compared to RGB cameras, RGB-D sensors may additionally accumulate errors introduced by the depth reconstruction process itself.

In the literature, many different methods for detecting objects made of transparent and reflective materials have been proposed. In [4], the detection of transparent objects using laser scans is based on the concept of Quasi-Static Environment Objects, i.e., objects whose detectability depends on the measurement location. In the case of transparent glass panes, this means that they are clearly visible only at beam incidence angles close to zero. In [5], changes in reflection intensity caused by specular reflection are exploited to detect glass objects. The authors defined parameters describing the intensity curve for glass materials, including the minimum reflection intensity threshold, reflection intensity variation, and the width of the reflection intensity peak, which were then used to classify objects as transparent or non-transparent.

Kim and Chung analyzed reflective characteristics of laser beams to classify the type of reflection as geometrically specular or diffuse and combined this information with distance measurements from the first and second reflection beams to detect transparent objects [6]. Meng et al. applied a ray-casting localization method to reduce localization errors in the vicinity of transparent objects [7]. Singh et al. proposed a Bayesian filter based on the fusion of laser and sonar measurements to reduce uncertainty caused by transparent objects [8]. Similar data fusion approaches using different sensor combinations were presented by Yang and Wang [9] and by Nagla et al. [10].

Learning-based methods have also been investigated. Jiang et al. used an artificial neural network incorporating information about distance, reflection intensity, and beam incidence angle to classify objects as glass or non-glass [11]. Mora et al. proposed a method for detecting transparent and reflective objects based on differences in the relationship between reflection intensity, beam incidence angle, and distance for various materials [12]. In [13], Koch et al. exploited information from the first and second single-beam echoes and fitted the measured data to material characteristics derived from the Phong reflection model [14], comparing averaged echo responses to determine whether an object is transparent or reflective. Tibebu et al. presented a method based on differences in the variance of measured distance and reflection intensity in point clouds obtained from standard and transparent objects [15]. The resulting point cloud is processed using filters that apply experimentally selected thresholds to classify points as glass or non-glass and detect glass object boundaries. Existing methods predominantly rely on peak reflection intensity values for materials exhibiting strong specular reflection in order to detect glass, shiny metal, or mirror-like surfaces.

In this paper, an original approach to the material classification of objects detected by a laser scanner (LiDAR) commonly used in autonomous mobile robot navigation systems is proposed. The method operates on a single-frame LiDAR scan composed of single-beam echoes and explicitly takes into account the presence of objects that reflect light to varying degrees, including transparent glass surfaces. By focusing on single-frame measurements and interpretable classification criteria, the proposed approach provides a lightweight and computationally efficient solution suitable for integration into real-time navigation systems.

2. Materials and Methods

2.1. SLAM Sensors

SLAM algorithms are typically based on data acquired from laser scanners, RGB cameras, or RGB-D (depth) cameras [16]. The popularity of laser scanners in mobile robotics results from their relatively high measurement accuracy combined with low computational requirements. In the context of detecting reflective and transparent materials for environment mapping, research focuses primarily on exploiting information related to the intensity of the reflected laser beam and the beam incidence angle. In some approaches, information from additional sensors, such as ultrasonic sensors, is also considered.

Distance measurement using a pulsed laser scanner is based on the time-of-flight (ToF) principle. A beam of laser light is emitted towards an object, and the time elapsed until the reflected beam returns to the sensor is measured. Based on this time, the distance between the sensor and the object is calculated [17]. In addition to distance information, modern laser scanners provide the reflection intensity $[eqn]$ , i.e., the strength of the return signal received by the sensor. This value is usually expressed as a normalized or dimensionless quantity, depending on the sensor manufacturer.

The relationship between the reflected intensity $[eqn]$ , the reflection coefficient $[eqn]$ , and the intensity of the incident laser beam $[eqn]$ can be expressed as

$$[eqn]$$

This relationship enables the extraction of additional information about the properties of the surface from which the laser beam is reflected.

In addition to the traveled distance, the strength of the return signal is influenced by the reflection and scattering properties of the surface. Since a single laser pulse may generate more than one response, the returning signal is filtered based on reflection intensity and response time. In such cases, the measurement with the highest reflection intensity is typically selected, while in multi-echo sensors the first and last reflections may also be recorded. Multiple responses can occur because the laser beam has a non-zero angular width and may be reflected from more than one object. Moreover, the sensor may register external electromagnetic radiation with a wavelength similar to that of the emitted laser beam (e.g., due to solar radiation or artificial light sources).

The amount of energy returned to the sensor depends on both the distance to the object and the physical properties of the surface, including its shape, material, color, and surface roughness. After reflection, the energy loss is determined by light transmittance, absorption, and reflection characteristics. Two main types of reflection can be distinguished. Specular reflection occurs on smooth surfaces, such as glass panes, where the reflected beam follows the law of reflection with respect to the incidence angle $[eqn]$ . Diffuse reflection occurs on rough surfaces, where the incident beam is scattered in multiple directions. For transparent glass, specular reflection is dominant and is reliably detected only at small incidence angles. At larger angles, the reflected beam intensity may become too low to be registered. The physical phenomena accompanying light reflection from transparent objects are illustrated in Figure 1.

RGB cameras (monocular cameras) are relatively inexpensive sensors but typically require significant computational resources for SLAM. Localization and mapping are performed by processing image data to extract and match visual features, such as lines and keypoints [18]. Objects made of transparent materials, such as glass, are largely invisible to RGB cameras because visible light is transmitted rather than reflected from such surfaces [19].

RGB-D cameras (depth camera) provide point cloud data similar to laser scanners, enriched with color information. These sensors commonly operate based on stereo vision using two RGB cameras, sometimes supported by additional devices such as infrared illuminators. RGB-D cameras are relatively inexpensive and require moderate computational resources. However, like laser scanners and RGB cameras, they exhibit difficulties in detecting transparent objects. In stereo-based systems, transparent materials cause similar problems as in RGB cameras. In systems using projected infrared patterns, depth estimation depends on the visibility of the pattern on the object surface, which in the case of transparent materials depends strongly on the incidence angle, similarly to LiDAR. Additionally, RGB-D cameras are more sensitive to ambient light and may accumulate errors due to the use of multiple cameras.

Other sensing solutions are also employed in SLAM systems. However, using them as the primary source of localization or mapping often reduces overall system accuracy, limits detection range, or requires modifications to the environment. Such solutions include ultrasonic sensors, artificial markers, and bumpers.

2.2. Material Type Identification

In this subsection, reflection intensity characteristics associated with different material types are analyzed. Based on these characteristics, a material identification method was developed using nearest neighbour and k-nearest neighbour machine learning approaches, with modifications aimed at achieving the highest possible identification accuracy.

Materials can be categorized according to the reflection intensity obtained from the emitted laser beam. This value depends on both the distance to the object and the physical properties of the material from which the beam is reflected. Key properties influencing reflection intensity include transparency, absorption coefficient, reflection coefficient, and beam incidence angle, as illustrated in Figure 1. These properties determine the reflection mechanism and directly affect the intensity recorded by the sensor. A commonly used classification of material types is based on the distance and incidence angle of the reflected beam [11,12,13].

This classification distinguishes three material categories:

• standard materials, for which diffuse reflection is dominant (e.g., walls painted with matte paint, cardboard, canvas curtains);
• reflective materials, characterized by a predominant contribution of specular reflection (e.g., polished metal or marble tiles);
• transparent materials, for which specular reflection dominates while most of the incident light is transmitted (e.g., glass windows or glass doors).

For experimental purposes, reflection intensity characteristics were collected for selected material types. The selected material samples included: (i) transparent glass panels, (ii) reflective metallic plates with polished surfaces, and (iii) standard diffuse materials such as painted walls and cardboard panels. These materials were chosen as representative examples of transparent, reflective, and standard material classes commonly encountered in indoor environments. The data describe the dependence of reflection intensity on beam incidence angle and distance to the object. Measurements were performed using an Ubuntu 20.04 (Focal) machine with ROS noetic on Hokuyo UTM-30LX multi-echo laser scanner (Hokuyo Automatic Co., Ltd., Osaka, Japan) on reference samples representing three material classes: white cardboard (standard), a polished steel plate (reflective), and a 7 mm thick glass pane (transparent).

The samples were positioned perpendicular to the sensor and moved over a distance range from 0.5 m to 5.6 m in steps of 0.1 m. Due to the constant sample size, the angular range of the scan decreased with increasing distance, resulting in a varying number of measurements per distance. For each distance, 1000 static scans were recorded. Each single frame scan consisted of multiple single beam echoes reflected from the sample, providing distance d and reflection intensity $[eqn]$ values.

All samples used in the experiments were planar surfaces mounted on a rigid support. As a result, variations in the recorded reflection intensity were primarily a function of the beam incidence angle and the distance between the sensor and the surface. The experimental setup ensured that the relative orientation between the LiDAR sensor and the sample was controlled and repeatable for all measurements.

For each single frame scan, the average reflection intensity $[eqn]$ was computed for each beam. Averaging the reflection intensity over multiple single beam echoes was performed to reduce the influence of measurement noise and local surface irregularities. Preliminary analysis showed that the standard deviation of the normalized reflection intensity within a single configuration was small relative to inter-material differences; therefore, the mean value was considered a reliable representative descriptor for further analysis.

The complete dataset was then divided into two disjoint subsets using stratified random sampling: 80% of the single frame scans were assigned to the training subset and 20% to the test subset. The partitioning was performed prior to classifier configuration and profile construction. All model configuration steps were carried out exclusively on the training subset, while the test subset remained unseen and was used only for final evaluation.

Material-dependent intensity profiles $[eqn]$ were constructed using only the training subset. Each profile consists of the recorded distance d, the normalized reflection intensity $[eqn]$ , and the beam incidence angle $[eqn]$ . The normalized intensity $[eqn]$ was computed according to

$$[eqn]$$

The truncation and scaling thresholds correspond to the effective dynamic range of the Hokuyo UTM-30LX sensor and were determined empirically based on preliminary measurements. When applying the method to a different LiDAR sensor, recalibration of these thresholds would be required to account for different intensity scaling characteristics.

The resulting training profiles $[eqn]$ and the independent test subset $[eqn]$ were combined into the dataset $[eqn]$ , which additionally contains information about the material class $[eqn]$ associated with each measurement.

The measurement results confirm distinct differences in reflection intensity characteristics between material types, as shown in Figure 2.

For standard materials, the intensity curve does not exhibit a pronounced peak and decreases gradually with increasing incidence angle, reflecting the dominance of diffuse reflection. Reflective materials exhibit a clear intensity peak at small incidence angles, with a rapid decrease at larger angles due to specular reflection. Transparent materials show the highest peak at near-zero incidence angles, followed by a rapid drop to near-zero intensity values, indicating a dominant specular reflection component.

For all material types, reflection intensity decreases with increasing distance from the sensor, with the highest values observed at $[eqn]$  m (Figure 3, Figure 4 and Figure 5).

The estimated reflection intensity corresponds to the normalized sensor-reported intensity after compensation for distance-related attenuation. The observed noise in Figure 4 and Figure 5 is primarily caused by increased sensitivity of the intensity normalization at larger incidence angles, where small geometric deviations lead to relatively large intensity fluctuations. This effect does not significantly influence the classification results, as the proposed method relies on comparative profiles rather than absolute intensity values.

The theoretical inverse-square relationship

$$[eqn]$$

approximates this behavior, where $[eqn]$ is the intensity received by the sensor, $[eqn]$ is the emitted intensity, and d is the distance to the object. Measurements below 0.5 m exhibit increased intensity but also larger distance measurement errors than specified by the manufacturer; therefore, single frame scans acquired at distances below 0.5 m were excluded from further analysis.

2.3. Proposed Approach to Material Type Identification

The goal of the proposed method is to identify the material type of an object scanned by a LiDAR sensor. Each object is classified into one of three material classes: standard, reflective, or transparent. This classification is possible because materials belonging to these classes exhibit distinct reflection intensity characteristics as a function of beam incidence angle.

Material-dependent intensity profiles developed for the three reference material types serve as patterns in the classification process. Since the classifier is intended for applications in which interpretability and deterministic decision rules are required, the method is designed to rely on a small number of explicitly defined features and distance-based decision rules, avoiding implicit parameter learning. For this reason, black-box models such as neural networks were not used. Instead, classical instance-based classifiers were considered. Although support vector machines (SVMs) are effective in high-dimensional spaces, their performance degrades in regions where data from different classes overlap, as outliers are omitted [20]. Therefore, k-nearest neighbour (k-NN) methods were selected.

The k-NN classifier determines the K nearest patterns to a given measurement based on a selected distance metric (e.g., Euclidean) and assigns the class according to a majority vote [21]. The choice of K and the distance metric directly influences classifier performance. In this work, the Python 3.8.10 and SciPy 1.5.4 implementation of the k-NN classifier was used. For the classification phase, the k-NN algorithm was implemented using a KD-Tree structure to optimize the nearest-neighbour search. Although initial exploratory tests evaluated a broad range of neighbours (up to $[eqn]$ ), the final analysis focused on k = 3, 5, and 7 to capture localized data patterns. A uniform weight function was employed, ensuring that each neighbour exerted equal influence on the final prediction. All proximity calculations were performed using the Euclidean distance metric.

Each data point in the dataset $[eqn]$ is represented by a feature vector consisting of the distance d, incidence angle $[eqn]$ , normalized reflection intensity $[eqn]$ , and the corresponding material class label M.

The incidence angle was computed by grouping measurement points into segments and estimating the surface orientation, as illustrated in Figure 6. The method combines geometric segmentation, split-and-merge line extraction, and local surface fitting to produce a refined representation of the scene. The algorithm operates directly on sensor_msgs/LaserScan ROS messages and outputs structured object segments enriched with incidence angle values.

The processing pipeline consists of four major stages:

1. Laser scan preprocessing and Cartesian projection—distance and angle arrays are converted from polar to Cartesian coordinates, and invalid measurements are removed.
2. Split-and-merge segmentation—a recursive boundary extraction procedure identifies dominant line segments that are later used to subdivide raw object candidates.
3. Object extraction from raw scan—consecutive valid single beam echoes are grouped into potential objects based on continuity constraints. Very small groups are discarded.
4. Incidence angle estimation—for each segment, a local linear model is fitted using polyfit method, and the incidence angle at each point is computed as the angle between the laser ray and the tangent to the fitted surface.

The laser scan segmentation and incidence angle estimation algorithm (Algorithm 1) works as follows. Given a laser single frame scan message containing an array of distances with corresponding intensities and angular increments, the algorithm first removes invalid measurements, i.e., distances out of the scanner detection range. Each remaining sample is stored as a tuple of distance, normalized intensity, and single beam echo angle.

Consecutive valid measurements are grouped using a continuity constraint based on the expected geometric separation given the angular increment, according to Equation (4), where $[eqn]$ m, determined based on LiDAR specifications. Only groups exceeding five samples are further processed.

$$[eqn]$$

In parallel, a global split-and-merge procedure is applied to the Cartesian projection of the entire single frame scan. This recursive routine identifies points that maximally deviate from the line connecting the first and last point of a subsequence. If the deviation exceeds a threshold (0.1 m in this implementation), the sequence is subdivided; otherwise, the best-fit line is represented by its endpoints.

Object candidates are then subdivided using the extracted line boundaries. For each final segment, a local coordinate axis is selected based on point distribution (switching axes if necessary to avoid vertical-line degeneracy). A first-order polynomial is fitted using polyfit method, serving as a local surface approximation. For each sample, a tangent line is constructed, and the incidence angle is computed as the angle between the laser single beam and the surface tangent. The angle is finally normalized to the range $[eqn]$ .

Initial experiments showed that the distance-based k-NN classifier (KNN-D) is capable of distinguishing material types, but its performance decreases in regions where reflection intensity curves intersect (Figure 2). To address this limitation, an additional feature—the gradient of normalized reflection intensity $[eqn]$ —was introduced. This feature enhances discrimination in regions where material-dependent intensity profiles overlap; its quantitative impact on classification performance is evaluated in Section 3. Algorithm 1: Laser Scan Segmentation and Incidence Angle Estimation

The normalized single beam echo intensity gradient $[eqn]$ is computed according to Equation (5). For interior measurements, it is defined as the absolute difference between neighbouring normalized intensity values. For boundary cases (first and last measurements), forward or backward differences are used to avoid undefined indices.

$$[eqn]$$

Changes in the gradient of normalized reflection intensity as a function of the beam incidence angle for different material types are shown in Figure 7.

Using the previously built $[eqn]$ dataset, a new dataset $[eqn]$ was created by extending the feature vector with the gradient value $[eqn]$ for a given distance d and incidence angle $[eqn]$ . The extended dataset $[eqn]$ was used to train the k-NN classifier (denoted as KNN-G). Since experimental results indicated that KNN-G did not consistently outperform KNN-D, further modifications were introduced.

First, the nearest neighbour (NN) algorithm was used instead of k-NN, assigning each measurement to the class of the closest pattern (Figure 8).

Second, a weighted distance metric was introduced. The final class assignment is obtained by selecting the material class associated with the minimum weighted distance, similarly to [22]. The optimal values of the weighting coefficients were determined exclusively using the training subset. A systematic grid search was performed over the predefined parameter range, and the configuration yielding the highest performance on the training data was selected. The independent test subset was not used during weight selection or classifier configuration and was reserved solely for final evaluation.

3. Results

The performance of the proposed NN-W classifier depends strongly on the choice of weights used in the weighted distance metric. To determine suitable values of the weights $[eqn]$ and $[eqn]$ , a systematic evaluation was conducted exclusively on the training subset collected during the creation of material-dependent intensity profiles. In consecutive trials, both weights were varied independently in the range from 0 to 10 with a step of 0.01, covering all combinations within this interval. From the full set of tested weight combinations, only those achieving high detection efficiency for particular material classes, as well as configurations relying exclusively on a single metric component, were selected for further analysis. After selecting the final configuration based on the training data, classification experiments were performed once on the independent test dataset using the KNN-D, KNN-G, and NN-W classifiers.

The quantitative results obtained for standard, reflective, and transparent materials are summarized in Table 1, Table 2, and Table 3, respectively. The classifier configuration is indicated next to its name, where K denotes the number of neighbours for KNN-based methods and the pair $[eqn]$ denotes the weight values used by the NN-W classifier. To enable a comprehensive comparison of classifier performance, the evaluation metrics $[eqn]$ , $[eqn]$ , $[eqn]$ , $[eqn]$ , and $[eqn]$ score were used [23].

The best overall performance in the classification of transparent and reflective materials was achieved by the NN-W classifier with metric weight values $[eqn]$ and $[eqn]$ . For standard material classification, the KNN-D (K = 3) classifier was slightly better according to the summary metrics, while NN-W remained competitive. It should also be noted that relying solely on gradient information (NN-W with $[eqn]$ and $[eqn]$ ) significantly reduces detection efficiency across all material classes, confirming that the gradient alone is not sufficient as a discriminative feature.

The error classification ratio represents the percentage of incorrectly classified single-beam echoes. It is calculated by dividing the number of erroneous detections by the total number of detections and multiplying the result by 100.

The higher efficiency of the NN-W classifier for standard and transparent materials (Figure 9 and Figure 10) confirms that the weighted metric mitigates misclassification at the intersection regions of material-dependent intensity profiles, which are clearly visible in Figure 2. In contrast, KNN-D (K = 3) (Figure 11 and Figure 12) and KNN-G (K = 3) (Figure 13 and Figure 14) tend to misclassify points precisely in those regions where the corresponding intensity curves intersect.

This effect is less clearly visible for reflective material, where NN-W assigns more points to incorrect classes (Figure 15). This behaviour is caused by large variations in normalized reflection intensity as a function of the beam incidence angle for reflective surfaces, which introduce disturbances in gradient-based features. The impact of such disturbances is especially evident for KNN-G (K = 3) (Figure 16). In this case, KNN-D (K = 3) (Figure 17) performs best among the compared methods, as it does not rely on changes in the gradient of normalized reflection intensity.

Overall, the NN-W classifier with metric weight values $[eqn]$ and $[eqn]$ achieved the best performance for recognizing the material type of objects scanned by the laser sensor, providing a reliable trade-off between correct detection of standard, reflective, and transparent materials.

4. Discussion

The proposed approach demonstrates that material type recognition based on a LiDAR single-frame scan composed of single-beam echoes can be achieved with high accuracy by explicitly exploiting the relationship between reflection intensity, beam incidence angle, and distance to the scanned object. Unlike methods relying solely on intensity peaks or distance information, the proposed classifier performs material discrimination using a weighted nearest-neighbour rule applied to distance, incidence angle, normalized reflection intensity, and the gradient of normalized intensity.

Previous approaches [11,12,13] report high performance for material classification; however, they do not explicitly address regions in which material intensity profiles overlap across different classes, which can lead to incorrect material assignments. The proposed approach improves discrimination in such cases by incorporating local intensity variation. In particular, it enables distinguishing between measurements that share similar absolute intensity values but exhibit different gradients, thereby disambiguating overlapping material-dependent curves.

The results indicate that incorporating the gradient of normalized reflection intensity can reduce misclassification in regions where material-dependent intensity profiles intersect, as observed for standard and transparent materials. This effect is particularly evident for transparent materials, where distance-based classifiers tend to misclassify measurements near zero incidence angles, as shown in Figure 2 and Figure 10. By balancing absolute intensity values with local intensity variations, the weighted nearest neighbour (NN-W) classifier reduces ambiguity at these intersection regions, resulting in improved robustness compared to both KNN-D and KNN-G approaches.

At the same time, the results indicate that the effectiveness of gradient-based features depends strongly on surface geometry. While the proposed method performs well for planar surfaces, such as glass panes, metal plates, or walls, it is less suitable for irregularly shaped objects. In such cases, rapid local changes in surface orientation introduce significant variations in reflection intensity that degrade the reliability of the gradient-based descriptors. This limitation highlights an inherent constraint of methods relying on local incidence angle estimation from sparse scan data.

From the perspective of autonomous navigation, the presented approach offers a practical enhancement to perception systems operating in indoor environments containing transparent and reflective structures. Integrating the proposed material classification method with SLAM algorithms may improve localization and mapping performance in the vicinity of glass and polished surfaces, where conventional laser-based methods are prone to failure. Moreover, material-aware mapping could support more informed motion planning and obstacle avoidance strategies.

Future research directions include extending the method to operate on locally constructed maps rather than individual scan segments, which may allow more robust material recognition for complex object geometries. Another promising direction is the application of the proposed approach to data acquired from multi-line 3D laser scanners, potentially enabling material classification in three-dimensional environments. Further work will also focus on improving scan segmentation methods and on detecting and mitigating disturbances caused by mirror-like reflections, which remain a challenging source of interference in laser-based perception systems.

5. Conclusions

This paper presented a method for material type classification based on the analysis of reflection intensity obtained from a single LiDAR single-frame scan composed of single-beam echoes. The proposed approach enables distinguishing between standard, reflective, and transparent materials by exploiting material-dependent intensity profiles defined as functions of distance and beam incidence angle.

The classification method relies on a nearest neighbour scheme combined with a weighted distance metric that incorporates both the normalized reflection intensity and its local gradient. This formulation reduces ambiguity in regions where intensity profiles of different material classes intersect, which is a known limitation of purely distance-based classifiers.

The proposed approach operates directly on individual scans and uses a limited set of physically motivated features, without requiring temporal accumulation of data, multi-sensor fusion, or complex model training procedures. As a result, the method is suitable for real-time material recognition tasks in mobile robotic systems equipped with standard 2D LiDAR sensors.

Experimental results demonstrate that incorporating gradient-based information reduces misclassification for transparent and standard materials in regions where material-dependent intensity profiles intersect, while also highlighting limitations related to surface geometry and measurement noise. In particular, the effectiveness of gradient-based features decreases for reflective materials with rapidly varying intensity responses.

Future work will focus on integrating the proposed method with SLAM pipelines, extending it to 3D LiDAR data, and improving robustness for irregular surface geometries. These extensions may further enhance perception capabilities in autonomous navigation systems operating in environments containing reflective and transparent objects.