A Survey on Hough Transform, Theory, Techniques and Applications

• The Hough transform converts global image-space detection into local peak detection in parameter space.Source: §1 · p. 1
• Edge points on a line vote for parameter combinations, reinforcing the accumulator cell for that line’s slope and intercept.Source: §1 · p. 1
• Because slope-intercept form degenerates for vertical or near-vertical lines, the paper introduces the normal parameterization in, using distance and angle.Source: §1 · pp. 1–2
• Independent edge-point voting supports parallel processing and noisy or partly deformed shapes, while storage and computation rise rapidly with problem dimensionality.Prior knowledge can improve efficiency by reducing accumulator size.Source: §1 · p. 2
• The survey updates earlier comparisons and, covering HT variations, direct gray- and color-image use, efficiency remedies, and applications.Its application coverage highlights up-to-date and unconventional cases.Source: §1 · p. 2

2. Shape Detection

• The standard Hough transform (SHT) focuses on lines, while variations extend detection to analytical and irregular shapes.Lines are also used in applications ranging from indoor industrial settings to outdoor vehicle surveillance.Source: §2 · p. 2, §2.1 · p. 2

2.1 Line Detection

• Line-detection variants target accuracy, efficiency, noise robustness, or specialized tasks through windows, probabilistic voting, active gridding, and vanishing-point methods.The grid Hough transform replaces random point selection with active gridding; the cascaded transform also addresses vanishing-point mappings and checkerboard detection.Source: §2.1 · pp. 2–3

2.2 Other Analytical Shapes

• The circular Hough transform (CHT) detects circles through a three-parameter voting scheme, requiring more computation and storage than line detection; ellipse detection uses a five-dimensional parameter space.Circle methods include center detection followed by radius-histogram validation, while ellipse methods include symmetry-based approaches and iterative randomized detection of incomplete ellipses under high noise.Source: §2.2 · pp. 3–5

2.3 Irregular Shape Detection

• The generalized Hough transform (GHT) uses a look-up table linking boundary positions and orientations to Hough parameters, supporting irregular shapes and hierarchical construction of complex shapes.It is robust to occlusion and additional structures, but has high storage and computation demands.Source: §2.3 · p. 5

3. Tone and Color

• The section addresses information lost when gray images are converted to binary and surveys approaches that use richer gray or color information, including a direct-gray-image case study.The reviewed approaches range from adding gray values to accumulator arrays to color segmentation and color histograms.Source: §3 · p. 6

3.1 Irregular Shape Detection

• Gray- and color-image approaches include adding gray values to the accumulator, extracting homogeneous segments or regions, and using color similarity or spatial color histograms.Color-region extraction may miss objects composed of many small uniform regions, and its Hough-space quantization levels require careful choice.Source: §3.1 · pp. 6–7

3.2 Case Study: the Gradient Hough Transform

• The GrdHT works directly on gray images by using gradient magnitudes to exclude low-gradient regions from voting and a sensitivity parameter to tune detection.It is described as fast and able to detect line-segment endpoints.Source: §3.2 · p. 7
• The gradient Hough transform (GrdHT) detects many lines missed by the SHT in some images, but performs worse in noisy imagery and can overestimate lines in textured, noisy scenes.In one checkerboard image, GrdHT detects most lines while SHT detects only some horizontal lines; another image yields poor performance for both.Source: Figure 10 · p. 7, §3.2 · p. 7
• GrdHT is reported as more accurate than SHT in many cases, but its sensitivity setting is critical and image-dependent.The figure varies sensitivity from 0.02 to 0.6 and shows its effect on detected-line counts.Source: Figure 11 · p. 8, §3.2 · p. 9

4. Speedup and Memory Saving

• FPGA implementations store Hough parameters or accumulator data and exploit angle-level and pixel-level parallelism, with run-length encoding also used to skip computations and memory accesses.The survey describes FPGA implementations for line and circle detection as well as real-time line detection in high-definition video.Source: §4 · p. 9
• Parallel CHT implementations achieved maximum speedups of 3.648x on a 4-core system and 45.7x on a GPU, while another reported up to 65 times speedup over sequential execution.The first figures are from; the latter is from.Source: §4 · p. 9
• Voting acceleration includes PPHT’s use of only a fraction of points, while other methods target clutter discrimination, rotation and scale handling, or multiple-peak search.The cited approaches include a local parameter-space operator and global threshold, FGHT, GA with HT, and greedy inference.Source: §4 · p. 9
• Hough forests combine random forests with GHT for real-time object detection and tracking, while HT voting is also used for human action classification.Random trees learn mappings from densely sampled feature patches to votes in Hough space.Source: §4 · p. 9
• Other computational strategies include centroid-focused license-plate skew correction, gradient-based RHT mapping for crop rows, and GHT lookup-table indexing with quadratic dependence on the number of samples.These approaches are described in,, and, respectively.Source: §4 · p. 9
• As image quality worsens, EPSOHT achieves the highest accuracy in the lowest time among EPSOHT, IRHT, and RHT, whereas the approaches are close under low noise.The comparison is reported in.Source: §4 · p. 9

5. Hough Transform Applications

• HT applications span 3D work, object and shape detection, lane and road-sign recognition, industrial and medical uses, pipe and cable inspection, and underwater tracking.Source: §5 · p. 10

5.1 Traffic and Transport Applications

• Lane-detection approaches combine HT variants with hierarchy, line partitioning, gradients, JPEG compression, inverse perspective mapping, or split-image processing.The survey notes that the JPEG-combined approach was tested only in simulations; the IPM approach assumes a flat road and uses camera calibration and pose information.Source: §5.1 · p. 10
• Transport applications also include road extraction from remote-sensing and SAR images, license-plate detection and skew correction, road-sign recognition, and railroad defect inspection.The SAR method uses an average HT on line intensities; the railroad example uses CHT to detect missing bolts, defects, and cracks in ties.Source: §5.1 · pp. 10–11

5.2 Biometrics and Man-Machine Interaction

• Hand applications use PHT for line features in an articulated hand model, GHT for Turkish finger-spelling recognition, and HT with connected components for real-time gesture interaction.The interaction method extracts the hand center, directions, and outstretched-finger positions; these examples are reported in,, and.Source: §5.2 · p. 11
• Fingerprint matching uses a descriptor-based HT alignment method and compares fingerprints using both minutiae and orientation-field information.Source: §5.2 · p. 11
• Eye-related applications use CHT for eye and head localization, and combine it with image statistics, adaptive thresholding, active contours, or radial gradients for pupil and iris processing.For iris segmentation, CHT provides an outer-boundary estimate and an initial pupil contour for active-contour segmentation; the outer boundary is circularly approximated while the inner boundary follows its shape.Source: §5.2 · p. 11

5.3 3D Applications

• For 3D reconstruction, HT methods detect lines, boxes, blocks, planar faces, and roof planes, while vanishing-point detection is identified as valuable for reconstruction.The 3D detector in uses 2D Hough space for parallel-line detection to reduce computational cost.Source: §5.3 · p. 11
• For roof-plane detection, the cited comparison reports RANSAC as faster and more efficient than HT, while HT is more sensitive to segmentation parameters and RANSAC can generate nonexistent constructions.A proposed extension harmonizes the algorithm with roof geometry.Source: §5.3 · p. 11

5.4 Object Recognition

• Object-recognition methods combine GHT with color similarity or use Hough-space voting to group features, with examples addressing illumination change, occlusion, distortion, clutter, and multiple object instances.The approaches are reported in,, and.Source: §5.4 · pp. 11–12
• The survey also describes HT-based methods for complex or broken patterns and shape retrieval, including region-level voting that works with gaps, rotations, scale changes, and shifts.The methods include PLHT and FIHT, EN-GHT for multi-break patterns, and region-level voting for arbitrary-shape retrieval.Source: §5.4 · p. 12

5.5 Object Tracking

• Radar surveillance uses the HT to initiate tracks amid many objects and noisy measurements of uncertain origin; another method filters clutter with the SHT and follows candidate tracks using the RHT and minimum-distance metric.The latter finds actual tracks using subtractive clustering.Source: §5.5 · pp. 12–13
• A mixture of uniform and Gaussian Hough (MOUGH) locates and tracks objects against complex backgrounds such as dense foliage while the camera moves.Source: §5.5 · p. 12
• For sparse range-image sequences, the bounded Hough transform (BHT), a GHT variation, exploits coherence across frames for object tracking.Source: §5.5 · p. 12
• Rectangular-object tracking applies the HT to Canny edges inside a predefined window, while a spatial color histogram with a GHT location estimator tracks objects between frames,.Source: §5.5 · p. 12
• Planar pose estimation compares numerical Hough-space representations with test patterns, while indoor robot motion estimation uses the HT to match tracked lines to walls.Source: §5.5 · p. 12

5.6 Underwater Applications

• An underwater-robot shape-recognition method formulates the problem as bounded-error estimation rather than using the classical GHT.Source: §5.6 · p. 13
• For underwater infrastructure, the HT detects pipelines,, tracks cables, and estimates pipeline boundaries and location in complex marine environments.The approach uses PHT to find line segments and Hough pruning to merge semicollinear lines; parameterizes pipeline boundaries as curves.Source: §5.6 · p. 13
• A micro-ROV uses video frames and PPHT to test fish-cage net integrity, while a subsea system detects and tracks man-made objects in poor visibility and marine snow.Source: §5.6 · p. 13

5.7 Industrial and Commercial Applications

• A UAV power-line inspection method applies a PCNN filter before the LHT and refines detections through knowledge-based line clustering; another HT variation classifies inspected items as pass or fail.Source: §5.7 · p. 13
• Industrial examples include HT-based inspection of clamps, holes, and welds, CHT-based pipe-joint detection, and combining CHT and LHT to detect natural seepage alterations.A CHT cell-detection algorithm approximates bee-comb cell spaces as circles; HT-based barcode line detection supports skew correction.Source: §5.7 · p. 13

5.8 Medical Applications

• Medical applications use the HT for ultrasonic defect localization, vertebra detection with a genetic algorithm [68], and surgical-instrument tracking with a particle filter.For instrument tracking, the HT is described as robust to illumination changes, occlusion, and distractions; PCA-derived gradient directions improve edge orientation accuracy and HT computation speed.Source: §5.8 · p. 13

5.9 Unconventional Applications

• A data-mining approach maps data space into parameter space to find arbitrarily oriented subspace clusters, including clusters of different dimensionality in noisy environments.Source: §5.9 · p. 14
• Feature matching is framed as density estimation in homography-hypothesis Hough space; the survey also notes an HT step in SIFT and scene classification using structure and lines.Source: §5.9 · p. 14

6. Conclusions

• HT variations span line detection through irregular shapes, with further variants expected to approach recognition of more complex objects.The authors attribute the transform’s research appeal to noise immunity, occlusion handling, and expandability.Source: §6 · p. 14
• Work has begun applying the transform directly to gray and color images, and the authors expect more color-image research that retains more information.Source: §6 · p. 14
• Applications span traffic, biometrics, object recognition and tracking, medical, industrial, and commercial fields, with unconventional uses still possible.Source: §6 · p. 14
• The authors expect memory-efficient, faster implementations and GPU parallel processing to support more time-critical applications.Source: §6 · p. 14