Feature detection refers to computing abstractions from the more delicate details extracted from an image and drawing accurate decisions at every point to discern the existence of an image element of a given kind at any given location. Ideally, feature detection marks the beginning point for milestone removal and association of data. It bears a close resemblance with interest point detection. Interest point detection, in the same way, aims at extracting local information content to aid in the summary of the vision system processing. The basis of its classification revolves around four significant divisions that are inclusive of the edge, corner, ridges, and blobs. The following are some of the techniques used in feature detection;
Interest point detection
It’s a common technique in visual SLAM. Herein, the distinctive point features comprise vivid mathematical procedures together with a better-outlined stance within the region covered by the image. Interest point detection avails rich information that summarizes advanced preparation in the vision system (Detone, Malisiewicz &Rabinavich, 2018). Unlike defamation, it’s less prone to disturbance. Interest point detection combines corner and blob detection since they bear similar characteristics. The only distinctive element that brings out the difference occurs in cases where the image is small. The corner detection focuses on the sharp image characteristics while the blobs majors on bringing out the smooth image characteristics.
By boosting Morava’s corner detector, this technique has implemented a combination of corner and edge detectors to examine image elements with fineness and random characteristics. According to their research, for the efficient performance of their technique, the image element must be discrete and should never be a continuum at the texture or edge pixels. They argued that the method was designed since curved lines and center edges can be fragmented differently (Usmani et al., 2021). The pros include its reliability, low numerical complexity, and invariance to image change, rotation, and illumination.
The only drawback of this technique is that its scale doesn’t change when multiplied by a common factor.
Harrice –Laplace detector
To cater to the invariant element feature, they introduced the Harrice-Laplace detector. It was designed majorly to tackle the distinctive affine transformations. An affine point is said to embed with its countable joints to countable joints with equidistant lines to equidistant lines. Still, the interconnection amidst the joints and angle formed by the lines are susceptible to change. The Harris detectors were used as an equalizer with a visual landmark in the independent blimp scheme of LAAS-CNRS.The pros of the Harris detector are its stable characteristic. In addition, its alterative nature proves its reliability in examining similar interest points under different views.
Shi and Tomasi detection
This detection technique focuses on viewing an image’s quality aspects. This technique employs striking differences that measure the change in the outlook of an element that appears in the first and current frame. This detection technique also provides evidence that authentic interpretation is an inadequate model for image movement when quantifying differences. Again, affine image alters; therefore, linear warping and interpretation are used sufficiently. Herein, the Newton Raphson minimization technique is used to resolve affine changes(Bansal et l,2021).In most instances, elements that contain high textures pose significant challenges with issues to do with poor illumination, depth discontinuity, occlusion, and drift away from their initial target. Features with good texture characteristics are selected to ensure the tracker works to its optimum.
Davidson and Murray (2002) applied the technique to recognize the regions of interest and matched related frames using the normalized sum of squared differences correlation.
Scale-invariant feature transform
David developed this technique in 1999. It came as an advancement in technology to improve the earlier versions of the transformation of images. It converts an idea into a comprehensive collection of local feature vectors. In addition, it involves the computation of a two-dimensional Gaussian function. It achieves this performance by combining the input image with the movement of a one-dimensional process in the horizontal and vertical directions. Lowe used the maxima and minima points of a Gaussian function to generate one characteristic vector that connotes an image to be applied as a parameter.
In addition, several researchers have invested a lot in research to advance the SIFT algorithm for the SLAM problem. The so-called autonomous semantic map construction constitutes one instance where the algorithm has been implemented. The HARI-SIFT-KAL technique has resulted in a move to end localization and kidnapped robot challenge.
Speeded-up Robust Features (SURF)
According to Bay et al. (2008), the SURF approach bases its foundation on the Hessian matrix and the collection of a two-dimensional Haar wavelet response. Ideally, SURF is considered vigorous. Again, they depend on the Hessian’s determinant for the location and scale selection. Generally, it was selected over SIFT because of its speed. In addition, its feature is rotation and does not change when sales are multiplied by a common factor.
Features from accelerated segment test
This technique was developed by Rosten and Drummond (2005, 2006). Its improvement resulted from a combination of edge and point-based tracking systems to counter the challenge of a real-time three-dimensional model-based tracking system. The table below points out the advantages and disadvantages of the point and line features.
Features from accelerated segment test
These features were brought forth by Rosten and Drummond (2005,2006). They combine the different elements of edge and point-based tracking systems to resolve the challenge of a real-time three-dimensional model-based tracking system. The edge and point based are vigorous when modeled together since it drives the system to its optimum performance.
This technique is designed to examine and figure out the distinct point in a digital image where the image brightness alternates brightly. The change in the element characteristics can be attributed to a wide array of factors such as discontinuities in depth and surface or could be triggered by variation in radiance and alteration in the component property. In most instances, interest point techniques are preferred by SLAM researchers. Usually, edge detection techniques are selected to cover up the drawbacks (Znamenskaya &Doroshchenko,2021). The edge detectors are grouped based on search-based and zero crossing-based categories.
Canny edge detector
It is a well-localized edge landmark and a good step-by-step process for choosing the landmark that exists. It’s usually selected since it contains high-order geometric details that are vital during and after SLAM. The study focused on estimating edges of identified regions and orientation. (Asmar et al.) also used Cranny’s step-by-step procedure to estimate distinct edges in the vertical direction of a tree trunk (Sangeetha & Deepa,2019).In this scenario, the tree is a landmark for outdoor environmental structures. Canny’s algorithm was advanced by several scholars to optimize the chances of finding a tree by adding more weight to the vertical. Canny’s edge detector produces very thin and clean edges in comparison to other technologies proposed by other scholars. The most pronounced drawback is that it consumes a lot of time due to its advanced and complex procedures involved in computation.
Sobel edge detector
This detector has been utilised by Shaw and Barnes (2006).They implemented a detector that aim to locate rectangular shaped objects that would bring out the edge joints that are located along disappearing lines. They otherwise blurs edge components which in turn brings forth distinct line to be appear more detectable. According to Harati et al (2007) planar surface are more conducive to use in comparison to point cloud raw data since points are no longer useful to applied in mapping for an inside environment(Xiang et al,2018).They applied the one dimensional measure called the Bearing Angle. It’s again preferable to use edge based techniques and BA-based division in navigation. However, in the use of FAT and Sobel algorithm, FAST is applicable in the identification of corner properties while Sobel is a technique suitable for conducting test on possible lines.
Vision slam with a system-on-chip approach
This technique combines myriad systems on a single silicon chip. Most developers and scholars in recent years have shifted to special purpose hardware to escalate image processing computation. In addition, the technology promises lighter weight, it’s cheaper and it consumes less power on average. The onboard solution again is desired to alleviate latency challenges that futures in the act of conveying a video.
The study has stipulated myriad techniques with the advancement that has been improved by various developers and scholars across the globe. This research plays a crucial role as far as detecting credible solutions to SLAM problems is concerned. Despite the popularity of interest point detection forming an integral part of most research work, it has been noted that edge and point detection has emerged credibly feasible in the present research. The technology in the present is seen to have advanced with the introduction of the system on chip technique. This technology is seen to turn the flow of events in another direction as robustness and real-time is optimized on another level altogether. In order for marketing to be done in real-time, FPGA technology is considered.
DeTone, D., Malisiewicz, T., & Rabinovich, A. (2018). Super point: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops (pp. 224-236).
Usmani, S., Saboor, A., Haris, M., Khan, M. A., & Park, H. (2021). Latest research trends in fall detection and prevention using machine learning: a systematic review. Sensors, 21(15), 5134.
Bansal, M., Kumar, M., Kumar, M., & Kumar, K. (2021). An efficient technique for object recognition using Shi-Tomasi corner detection algorithm. Soft Computing, 25(6), 4423-4432.
Znamenskaya, I., & Doroshchenko, I. A. (2021). Edge detection and machine learning for automatic flow structures detection and tracking on schlieren and shadowgraph images. Journal of Flow Visualization and Image Processing, 28(4).
Xiangxi, Z., Yonghui, Z., Shuaiyan, Z., & Jian, Z. (2018, October). FPGA implementation of edge detection for Sobel operator in eight directions. In 2018 IEEE Asia Pacific Conference on Circuits and Systems (APCCAS) (pp. 520-523). IEEE.