Visual impairment (VI) is a significant public health concern that affects people of all ages and is caused by a range of factors, including age-related eye diseases, genetic disorders, injuries, and infections [1]. The global population of individuals suffering from VI, including those who are completely blind, moderately visually impaired, and severely visually impaired, has reached more than 300 million [2]. The increasing number of visual impairment (VI) cases highlights the critical need to improve accessibility and mobility for visually impaired individuals, who face significant challenges in navigating public spaces due to the low success rate of obstacle avoidance.

Therefore, governments of different countries are attempting to design various assistive living facilities for individuals with visual impairments. In the United States, guide dogs and white canes remain essential tools. In addition, the emergence of advanced technologies has also enhanced the independent mobility of individuals with visual impairments and blindness. GPS-based navigation systems, such as smartphone applications and standalone devices, provide step-by-step navigation and information about points of interest. Furthermore, obstacle detection devices and electronic travel aids, such as ultrasonic canes and wearable sensors, assist individuals in navigating their surroundings [3]. In the United Kingdom, tactile pavements and signage have been implemented in public spaces to improve accessibility and orientation [4]. The “Haptic Radar” system in Japan utilizes vibrations to provide real-time feedback on surrounding objects [5].

However, the accessibility of these facilities is often inadequate in older districts, leading to the use of personal navigation tools such as white canes and guide dogs [6]. While white canes are a popular option, their short range and potential interference with other pedestrians may hinder mobility in crowded spaces. Alternatively, guide dogs offer effective guidance, but their high cost and restrictions on public transportation may limit their widespread use [7]. For the existing advanced technologies, engineers and manufacturers face technical challenges in ensuring the accuracy and reliability of navigation and object detection systems [3]. In daily life, it is essential to prioritize efforts to address the challenges faced by visually impaired individuals, as the loss of eyesight can be a debilitating experience.

The camera is a widely used and cost-effective sensor for environmental perception. It mimics the capabilities of the human visual system, excelling in the recognition of shapes and colors of objects. However, it does have limitations, particularly in adverse weather conditions with reduced visibility.

The radar (radio detection and ranging) is widely used to precisely track the distance, angle, or velocity of objects. Radars can be broken down into a transmitter and receiver. The transmitter sends radio waves in the targeted direction and the waves are reflected when they reach a significant object. The receiver picks up the reflected waves and gives information about the object’s location and speed. The greatest advantage of the radar is that it is not affected by visibility, lighting, and noise in the environment. However, compared to a camera, a radar is low-definition modeling and is weak at providing the precise shape of objects and identifying what the object is.

The mechanism of the LiDAR (light detection and ranging) is similar to the radar but utilize laser light to determine ranges instead of radio wave. The LiDAR is a more advanced version of a radar that can provide extremely low error distance measurement. It is also capable of measuring thousands of points at the same time to model up a precise 3D depiction of an object or surrounding environment [16]. The disadvantages of the LiDAR are its high cost and the requirement of a remarkable amount of computing resources compared to cameras and radars.

Although the costs of cameras, radar systems, and LiDAR can vary significantly due to factors such as brand, specifications, and quality, a general assessment of equipment costs with comparable capabilities reveals the following: Cameras typically range in price from $100 to several thousand dollars, depending on factors such as resolution, image quality, and additional features. Radar systems used for object detection and tracking start at a few hundred dollars for basic short-range sensors, while more advanced and specialized radar systems can cost several thousand dollars or more. Likewise, LiDAR sensors range in price from a few hundred dollars for entry-level sensors to several thousand dollars for high-end models with extended range, higher resolution, and faster scanning capabilities.

Considering the pros and cons of the three types of sensors for distance measurement, the camera is the most appropriate sensor to be utilized in the research due to its low cost, being less sophisticated, and its high definition. The 2D information recognized by the camera can be adopted directly by the deep learning algorithms of object detection.

Typical photos taken from a monocular camera are shown in two dimensions that would require extra information for distance estimation. Distance estimation (also known as depth estimation) is an inverse problem [17] that tries to measure the distance between 3D objects from insufficient information provided in the 2D view.

The earliest algorithms for depth estimation were developed based on stereo vision. Researchers utilize geometry to constrain and replicate the idea of stereopsis mathematically. Scharstein and Szeliski [18] conducted a comparative evaluation of the best-performing stereo algorithms at that time. Meanwhile, Stein et al. [19] developed methods to estimate the distance from a monocular camera. They investigated the possibility of performing distance control to an accuracy level sufficient for an Adaptive Cruise Control system. A single camera is installed in a vehicle using the laws of perspective to estimate the distance based on a constrained environment: the camera is at a known height from a planar surface in the near distance and the objects of interest (the other vehicles) lie on that plane. A radar is equipped for obtaining the ground truth. The results show that both distance and relative velocity can be estimated from a single camera and the actual error lies mostly within the theoretical bounds. Park et al. [20] also proposed a distance estimation method for vision-based forward collision warning systems with a monocular camera. The system estimates the virtual horizon from information on the size and position of vehicles in the image, which is obtained by an object detection algorithm and calculates the distance from vehicle position in the image with the virtual horizon even when the road inclination varies continuously or lane markings are not seen. To enable the distance estimation in vehicles, Tram and Yoo [21] also proposed a system to determine the distance between two vehicles using two low-resolution cameras and one of the vehicle’s rear LED lights. Since the poses of the two cameras are pre-determined, the distances between the LED and the cameras, as well as the vehicle-to-vehicle distance can be calculated based on the pinhole model of the camera as the focal lengths of the cameras are known. The research also proposes a resolution compensation method to reduce the estimation error by a low-resolution camera. Moreover, Chen et al. [22] proposed an integrated system that combines vehicle detection, lane detection, and vehicle distance estimation. The proposed algorithm does not require calibrating the camera or measuring the camera pose in advance as they estimate the focal length from three vanishing points and utilize lane markers with the associated 3D constraint to estimate the camera pose. The SVM with Radial Basis Function (RBF) kernel is chosen to be the classifier of vehicle detection and Canny edge detection and Hough transform are employed for the lane detection.

Nowadays, to achieve depth estimation using a monocular camera, neural networks are commonly used. Eigen et al. [23] proposed one of the typical solutions that presented a solution to measure depth relations by employing two deep network stacks: one that makes a coarse global prediction based on the entire image, and another that refines the prediction locally. By applying the raw datasets (NYU Depth and KITTI) as large sources of training data, the method matches detailed depth boundaries without the need for superpixelation.

Another solution that can overcome the weakness of using CNN for depth estimation is that vast amounts of data need to be manually labeled before training [24]. A CNN for single-view depth estimation that can be trained end-to-end, unsupervised, using data captured by a stereo camera without requiring a pre-training stage or annotated ground-truth depths. To achieve that, an inverse warp of the target image is generated using the predicted depth and known inter-view displacement to reconstruct the source image; the photometric error in the reconstruction is the reconstruction loss for the encoder. Zhou et al. [25] also presented an unsupervised learning framework for the task of monocular depth and camera motion estimation from unstructured video sequences. The system is trained on unlabeled videos and yet performs comparably with approaches that require ground-truth depth or pose for training. As a whole, Table 1 highlights the various deep-learning-based approaches to depth estimation.

Existing Solution	Technique/Model	Hardware	Target Object	Advantages	Disadvantages
Ye and Qian [26]	3D point cloud	3D ToF camera & tablet	Structural Object (e.g., doorway, hallway, stairway, ground, and wall)	Highly accurate and possible to combine with SLAM/wayfinding solutions	High computing resources required and indoor only
Kayukawa et al. [27]	YOLOv3-tiny	Smartphone (built-in RGB camera and infrared depth camera)	Human	High mobility and off-the-shelf device required	Very specific application and short distance
Ying et al. [28]	YOLOv3	Stereo webcam & NVIDIA Jetson TX2	Indoor furniture (e.g., chair and table)	Low cost but small dataset required	Low mobility and accuracy in distance estimation
Shelton and Ogunfunmi [29]	AlexNet	Webcam & laptop	Indoor objects and outdoor buildings	Text-to-speech function involved, available in both indoor and outdoor	Only workable in the authors’ campus and low mobility
Ryan et al. [30]	MobileNet-SSDv2	Micro-controllers, Raspberry PiCam & webcam, ultrasonic & infrared ToF Sensor	General objects (VOC and COCO dataset)	High mobility, available in low power, and low cost	Additional sensors for distance estimation required. Implemented with the existing navigation tool
Sohl-Dickstein et al. [31]	Ultrasonic echolocation	Speaker & ultrasonic microphones	Any object in short distance	Work without visible light, provide 3D spatial information	Short distance and could not recognize objects

An experiment (Chou, K.S.; Wong, T.L.; Wong, K.L.; Shen, L.; Aguiari, D.; Tse, R.; Tang, S.-K.; Pau, G. A Lightweight Robust Distance Estimation Method for Navigation Aiding in Unsupervised Environment Using Monocular Camera. Appl. Sci. 2023, 13, 11038. https://doi.org/10.3390/app131911038 ) shows that NFD provides a satisfying result in detecting selected near-front objects and estimating their distance from the user. It provides a relatively affordable solution for visually impaired people with the concept of ‘grab, wear, and go’. NFD utilized YOLOv4-tiny for object detection. It provides competitive performance in terms of accuracy among the other solutions for object detection. The training and inference speed outperform other solutions. However, the accuracy of distance estimation of NFD is not better than those with depth sensors, such as ToF cameras and LiDAR. However, inference output can directly locate detected objects in front at comparatively fewer resources (without going through the point cloud). On the other hand, using public objects, NFD can generally work in the entire city, whereas most of the existing solutions could only be utilized for indoor environments or particular areas. However, it can only recognize two types of outdoor public objects currently, implying that it can only work outdoors. Also, moving objects, such as humans and cars, are not detectable yet. To extend it, enhancing the dataset of the trained model for the improvement of the solution is needed in the future. The more high-quality data included in the dataset, the more accurate the prediction it can make using deep learning.

While deep learning technology has showcased its proficiency in depth perception and measurement, certain challenges persist: (i) Specialized Equipment: Generating media data with depth information necessitates specialized equipment like Kinect cameras, ToF cameras, or LiDAR sensors to create training datasets. Without such equipment, the laborious task of manually labeling each object with ground truth distance becomes inevitable. (ii) Unsupervised Framework: Unsupervised monocular camera depth estimation typically relies on stereo video sequences as input. It leverages geometric disparities, photometric errors, or feature discrepancies between adjacent frames as self-supervised signals for model training.