Cracks in paved roads are one of the most potent indicators of pavement damage. Cracking in the pavement is quite unavoidable, and there are many underlying factors (e.g., exposure to the sun, rain erosion, natural weathering, and long-term driving of vehicles) that accelerate the cracking of the pavement’s surface. If these cracks cannot be localized and repaired in time, they will have a negative impact on the safe driving of vehicles. Consequently, it can cause deadly accidents, as well as expenditure of a huge amount of money for the maintenance and repair of pavements. Therefore, crack detection at an early stage is essential to maintain the structural integrity and serviceability of paved roads. In past decades, manual crack detection was a very common practice for localizing cracks on paved roads. However, the manual method lacks efficiency and accuracy; it is expensive because of the necessary expertise. Moreover, it is considerably tedious, arduous, and time-consuming because experts must monitor the cracks with the naked eye by roaming the roads. Therefore, to lessen the workload of experts and make the system fast and cost-effective, researchers are bringing automation to crack detection. With the advancement of computer vision (CV) technology, various vision-based methodologies have already been developed to perform automatic crack detection. Early implementation of the CV techniques for crack detection was to some extent limited to threshold-based approaches (e.g., pixel intensity was used as the feature) ^[1][2][2,3], and other hand-crafted feature-based approaches. Some of the prominent hand-crafted feature-extraction techniques are wavelet features ^[3][4], Local Binary Pattern (LBP) ^[4][5], Digital Image Co-relation ^[5][6], Gabor filters ^[6][7], and so on. But these methods can only extract local patterns instead of global patterns, which pulls the detection results backward. Some research ^[7][8][9][8,9,10] has used model-based, traditional CV algorithms, which use geometric characteristics of images to perform crack detection globally. The advantages of model-based techniques over feature-based techniques are that model-based techniques can detect cracks in adverse conditions such as noisy environments, poor illumination conditions, and shadow problems. Though these model-based methods can partially solve noise problems and can detect cracks more continuously, their performance is not satisfactory enough when detecting cracks with complex patterns or intensity inhomogeneity.

In recent years, Deep Learning has been extensively applied in CV tasks for its noteworthy representation ability. DL models do not need hand-crafted features; rather, they can extract valuable features (both local and global) automatically from the input data. A few research works have already devoted their efforts to utilizing the properties of deep learning mentioned above to learn robust feature representation and detect cracks with more precision. Zhang et al. introduced a Convolutional Neural Network (CNN) classifier for the first time in 2016 to detect cracks in concrete structures ^[10][11]. The primary objective of this study was to develop a patch-based classifier to detect cracks in concrete structures. Later on, Cha et al. ^[11][12] and Eisenbach et al. ^[12][13] also performed patch-based classification, which can only identify the presence or absence of cracks in a corresponding image patch. Researchers also utilized another deep learning scheme called object detection for localizing the cracks, along with identifying them in an image ^[13][14][14,15]. However, these models can only classify and localize the cracks in a concrete structure instead of detecting cracks at a pixel level. So to solve this issue, Yang et al. incorporated an image segmentation technique for detecting concrete cracks at pixel level ^[15][16]. Crack segmentation involves classifying each of the pixels in an image as ’crack’ or ’non-crack’. Instead of detecting the class only in an image, crack segmentation detects an output image, highlighting the pixels containing the cracks, which localizes the cracks and extracts the original shape of the cracks. Moreover, the segmented images can later be used for the important task of extracting length, width, and area of cracks, which provides information about crack severity in concrete structures. Considering the advantages of crack segmentation over crack detection and classification, researchers from all over the world are devoting their efforts to developing crack-segmentation methods and quantifying the cracks to present an automated crack-detection system ^[16][17][17,18].

In past decades, various researchers have developed robotic vehicles to automatically inspect cracks. The first study, dating back to 2007 ^[18][19], designed an automated inspection system for cracks in concrete tunnels using a mobile robot. They collected the images using a CCD camera, which was interfaced with the robot, and stored the images in the robot’s brain. Later, they extracted cracks on a different computer using the Sobel edge-detection algorithm. Oyekola et al. also designed a robotic system for detecting cracks on concrete tank surfaces ^[19][21]. The authors also first collected the images and later detected cracks using a thresholding algorithm developed using the MATLAB programming language. Li et al. utilized the robotic platform developed by Guimu Robot Co Ltd., Shanghai, China for detecting cracks on pavement structures ^[20][22]. The authors developed an unsupervised algorithm named the Multiscale Fusion Crack Detection (MFCD) for inspecting the cracks. However, in this research, the cracks are also not detected by the onboard computer. La et al. developed a wall-climbing robot for detecting cracks on steel bridges ^[21][23]. The robot was equipped with several sensors and a camera. Navigating through the steel bridges, it collected data and passed them in real-time to the ground station for further processing and for detecting cracks using a Hessian-matrix-based filter. In another work, La et al. used the Seekur mobile robot platform and modified it by installing several NDE sensors (e.g., GPR, USW, ER, IE) and a camera for the monitoring of a concrete bridge deck ^[22][24]. The authors collected the images and passed them to the remote computer to extract the cracks using a Gradient-based algorithm. They also presented the delamination maps of the cracking using NDE data. The robot could localize itself and maneuver automatically on the bridge deck. However, this robotic system needed multiple onboard computers for navigating and processing everything. In the studies ^{[18][19][20][21][22]}[19,21,22,23,24], the researchers relied on conventional methods for crack assessment, employing image-processing algorithms and threshold-based approaches. While these methods provided initial insights, their weaknesses lie in accurate detection as well as in the lack of post-processing techniques to obtain comprehensive geometric information about the cracks. Additionally, the reliance on external computers for crack detection introduced potential delays and limitations in real-time assessment.

Hendrik et al. developed a legged robot named ANYmal for inspecting the crack conditions in underground tunnels ^[23][25]. The researchers considered the tactile sensory system instead of the vision data because of the presence of noise, water, etc., on the surface. They collected signals from the footsteps of the robot and classified the crack conditions using the Support Vector Machine (SVM) algorithm. The authors classified five types of crack conditions, including good, satisfactory, fair, critical, and failure, to provide information about the severity. Le et al. developed a mobile robotic system for the in-line inspection of the pipes ^[24][26]. The authors integrated multiple sensors (e.g., LIDAR, optic sensors) on the robot and classified these combined sensory data using the SVM algorithm for detecting cracks in pipes. Lei et al. developed a low-cost unmanned aerial vehicle for inspecting cracks in concrete structures ^[25][20]. They collected images using their UAV and classified cracks using the SVM classifier running on the onboard computer. Pan et al. utilized low-altitude images collected from a UAV to detect cracks on asphalt pavements ^[26][27]. The researchers collected centimeter-level spatial resolution images and utilized a hybrid model (ANN+SVM) to inspect the cracks. In ^{[23][24][25][26]}[20,25,26,27], machine learning techniques were introduced for crack assessment, improving upon the conventional methods. By using legged robots, mobile robots, and UAVs, these studies leveraged tactile sensory systems, the fusion of camera data with other sensors, and image analysis models such as Support Vector Machine (SVM), Random Forest (RF), and Artificial Neural Network (ANN). The focus was on classifying cracks and achieving enhanced crack detection using machine learning algorithms. However, weaknesses persisted in terms of limited classification capabilities and the reliance on specific sensor data fusion. The studies focused on crack classification rather than providing detailed geometric information, and the machine learning models were not fully capable of accurately quantifying crack properties such as depth and length. Montero et al. developed a mobile robotic system for detecting cracks in concrete tunnels ^[27][28]. They designed the mobile robot with an adjustable crane and a robotic arm. The adjustable crane carried the vision sensor and the robotic arm, while the robotic arm carried an ultrasonic sensor. They designed the crane to be adjustable so that it could reach different heights and directions for collecting data accurately. They collected images and passed them to the host computer, which analyzed the images using CNN, and they also contacted the ultrasonic sensor with the tunnel wall to analyze the cracks. Li et al. developed a quadrotor flying robot for detecting cracks in concrete bridges and tunnels ^[28][29]. The authors focused on reconstructing 3D metrics to determine the location of the defects and severity information using a visual-inertial fusion approach. They proposed a novel Deep Learning model named AdaNet to detect cracks using their own crafted dataset named Concrete Structure Spalling and Cracking (CSSC). Gui et al. developed a robotic system using the ARIR robotic platform for detecting cracks on airport pavement ^[29][30]. They utilized one vision camera and a GPR sensor for collecting surface and subsurface data. The data were passed to an analysis center to process the collected data. They employed an intensity-based algorithm to detect cracks from images and a voting-based CNN to predict the GPR data. Finally, the authors stitched the collected data together to present a continuous grid for visualization. Ramalingam et al. developed a mobile robotic platform named Panthera for segmenting cracks and detecting garbage on the road ^[30][31]. The authors adopted SegNet for the segmenting task and a DCNN-based object detector for detecting garbage. Furthermore, they also utilized the Mobile Mapping System (MMS) for localizing the defects. He et al. developed an unmanned surface vessel (USV) for inspecting cracks on the bottom part of bridges or urban culverts ^[31][32]. The researchers installed both RGB cameras and LIDAR to collect information from the environment. The authors proposed a novel Deep Learning model named CenWholeNet for detecting cracks. The USV was controlled from a ground station module, where the LIDAR and video information was also transmitted from the robot’s brain (Intel NUc mini pc). In ^{[27][28][29][30][31]}[28,29,30,31,32], advancements were made in crack assessment with the integration of deep learning techniques. These studies utilized CCD cameras, UAVs, and mobile robots to collect images and employ deep-learning models such as CNN and Adanet. The emphasis was on detecting cracks and estimating crack location. However, limitations were observed in terms of limited crack parameter estimation and the need for additional computational resources to handle the complex deep-learning algorithms. Moreover, comprehensive datasets for training and evaluating deep learning models were not extensively developed.

Yang et al. developed a wall-climbing robot for detecting cracks and spalling on concrete structures ^[32][33]. The researchers collected data using an RGB-D camera and estimated the cracks on the images by utilizing a novel deep learning model named InspectionNet deployed in Intel Nuc Mini PC. They also developed a map-fusion module for their work to highlight the detected cracks. Yuan et al. developed a mobile robotic platform for detecting cracks in reinforced concrete structures ^[33][34]. The authors proposed a Mask-RCNN-based model for segmenting the cracks on the images collected from a stereo camera. They utilized an NVIDIA Jetson Xavier device to implement the edge computing technique and pass the predicted frames to the host computer through the WebSocket protocol. They designed a UI for successfully controlling the robot and collecting the frames. Another important feature of this research is that after quantifying the damages, the researchers presented information about the actual size in a 3D cloud point reconstruction of the inspected structures. In ^[32][33][33,34], the research focused on real-time crack assessment using robotic systems, incorporating RGB-D cameras, stereo cameras, and LIDAR sensors. The integration of deep learning models on onboard computers demonstrated promising results. However, weaknesses include the need for fine-tuning algorithms to handle variations in lighting conditions and challenges associated with accurately reconstructing actual crack sizes. Additionally, limited attention was given to automated crack quantification and severity mapping. Table 1 presents a summary of the robotic platforms for crack inspection. Although there have been numerous remarkable research works in the field of automatically detecting pavement cracks, there is still a vast scope for improving existing methods.