A method for automatically tracking significant line segments in a highway tunnel based on a vehicle-mounted video sequence

Abstract

The application discloses a kind of highway tunnel hole inside significant line automatic tracking method for vehicle-mounted video sequence.The two significant line that the present application is particularly concerned about is the left and right split boundary between the dark area of tunnel top and white sidewall.By automatically tracking the two significant lines, and taking the intersection of the two significant lines as a blanking point, the jitter interference of continuous estimation of blanking point in vehicle-mounted video can be reduced, the performance of intelligent detection of LED lamp missing can be effectively improved, and the availability of existing work can be greatly improved. Experimental results show that the tracking effect of the significant line in the highway tunnel hole is good by using the technical scheme of the present application, which meets the actual needs of reducing the estimation interference of the blanking point in the continuous vehicle-mounted video and improving the performance of geometric information recovery in the road tunnel. This provides a new technical means for efficient acceptance and routine maintenance of tunnel lighting system.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering safety protection and vehicle-mounted video intelligent processing technology. Specifically, it relates to an automatic tracking method for significant line segments inside highway tunnels based on vehicle-mounted video sequences, which can provide new technical means for the efficient acceptance and daily maintenance of tunnel lighting systems. Background Technology

[0002] Highway tunnel lighting systems are crucial infrastructure for ensuring the safety and comfort of drivers in highway tunnels. The proper functioning of tunnel LED lighting fixtures directly affects the quality of the tunnel lighting environment. However, due to outdated traditional techniques for the acceptance and routine inspection of tunnel LED lighting fixtures, high labor costs, heavy inspection workload, and low testing efficiency are inevitable. Vehicle-mounted video systems can quickly capture real-time images of the actual working status of highway tunnel LED lighting fixtures over a wide area, and automatically analyze them using computer video intelligent processing technology, effectively improving the efficiency of acceptance, operation, and maintenance of tunnel LED lighting fixtures.

[0003] In recent years, Xin Le et al. have proposed a series of technical inventions regarding robust segmentation and automatic detection of missing LED lighting fixtures in highway tunnels. Specifically, see their authorized patents 1 and 2, "A Robust Detection Method for LED Lighting Fixtures in Highway Tunnels" (Patent No. 201910443626.2) and 2, "An Automatic Detection Method for Missing LED Lighting Fixtures in Tunnels Based on Digital Imaging" (Patent No. 201910728483.X). The inventors' work in this area provides a new fundamental technical means for the efficient acceptance and routine maintenance of tunnel lighting systems.

[0004] Patent 1, authorized by the inventor, primarily analyzes the geometric structure information of the road environment inside tunnels and proposes a robust method for rapidly and accurately detecting LED lighting fixture areas. The purpose of this method is to overcome various high-brightness interferences in complex scenes within highway tunnels. This method is also a prelude and necessary step for Patent 2, authorized by the inventor. Patent 2, based on the detection results of LED lighting fixtures in tunnels, delves into the linear cross-ratio invariance constraint of the camera imaging model, derives the rules governing the quantitative relationships between LED lighting fixture image areas, and achieves in-depth analysis and automatic detection of missing LED lighting fixtures. Summary of the Invention

[0005] This invention proposes an automatic tracking method for salient line segments inside highway tunnels based on vehicle-mounted video sequences. This invention builds upon the inventor's previous two patented technologies, further exploring video analysis of LED lighting fixtures in highway tunnels. The detection and tracking of hidden points is a commonly used technique in traffic image processing, greatly aiding in inferring the geometric attributes of traffic scenes. Restoring the geometric information of highway tunnels helps in more accurately analyzing the working status of LED lighting fixtures. This has already been fully demonstrated in the inventor's previous two patented technologies. However, when considering vehicle-mounted video analysis, a problem inevitably exists: video jitter, especially when shooting with handheld devices. This invention focuses on two salient line segments: the two key left and right boundaries between the dark area at the tunnel ceiling and the white sidewalls. By automatically tracking these two salient line segments and using their intersection as the hidden point, the jitter interference from continuous estimation of hidden points in vehicle-mounted videos can be reduced, effectively improving the performance of intelligent detection of missing LED lighting fixtures, thereby significantly enhancing the usability of existing methods. To achieve the above objectives, and considering the significant video jitter problem, this invention employs a motion estimation method based on local invariant features for automatic tracking of salient line segments in vehicle-mounted videos. Visual feature tracking is generally already highly mature. In real-world scenarios, the robustness of this method is guaranteed even in small video clips containing only a few image frames.

[0006] An automatic tracking method for salient line segments inside highway tunnels based on vehicle-mounted video sequences, characterized by the following steps:

[0007] Step 1: Robust detection of salient line segments based on geometric structure analysis of the road environment inside the tunnel

[0008] This step is performed entirely in accordance with the inventor's patent application, "A Robust Testing Method for LED Lighting Fixtures in Highway Tunnels" (patent application number 201910443626.2). This step focuses on the following three-stage algorithm:

[0009] First, the LSD (Line Segment Detector) operator is used to extract a set of line segments H representing the geometric structure of the internal environment of the highway tunnel from the original image of the tunnel LED lighting. Here, s is a line segment in H, defined by its two endpoints a1 and a2.

[0010] s={a1,a2}={(x1,y1),(x2,y2)}

[0011] Where a1 = (x1, y1) and a2 = (x2, y2) represent the starting point and ending point of s (in pixel coordinates), respectively.

[0012] Secondly, to avoid the negative impact of interfering line segments, a simple filtering process is performed on the line segment set H to remove lateral parallel lines from the road space ahead and the rear of other vehicles, thereby extracting the effective set of parallel line segments L that converge at the vanishing point in front of the main vehicle. Based on this, the M-estimator Sampling and Consensus (MSAC) algorithm is used to achieve robust detection of the vanishing point v in front of the main vehicle.

[0013] Finally, in the image space, a new coordinate system XOY is constructed with the vanishing point v in front of the vehicle as the origin. Clearly, in the new coordinate system XOY, each valid parallel line segment s will fall into only one quadrant. Based on a series of simple conditions, the line segments in the set L of valid parallel line segments are further evaluated and filtered, thus enabling robust detection of the dark area and white area at the top of the tunnel. sidewall The two key dividing lines on the left and right sides of the wall (b) l and b r .

[0014] Step 2: Calculate the expanded rectangle for each parallel line segment in the set L of valid parallel line segments.

[0015] Given time t, set L t Any parallel line segment in the polygon, when extended outwards, easily forms an expanding rectangle with an equidistant outward magnification effect. This expanding rectangle is actually composed of a series of straight lines parallel to the sides of the known significant line segment and spaced d apart. For a general polygon, let l i and l i+1 It is any two adjacent edges that intersect at vertex p. i Draw two sides outward from the inner polygon, parallel to l. i and l i+1 The distance between the two lines is d. These two lines meet at vertex q. i They intersect. Therefore, q i It is equal to vertex p in a parallelogram i The sum of two adjacent edges a1 and a2

[0016] q i =p i +(a1+a2)

[0017] Here, the directions of a1 and a2 are the directions of the corresponding sides of the inner polygon, and they can be represented by the difference between the corresponding vertices of the inner polygon. Unit vectors with the same direction as a1 and a2 can be represented as a1 / |a1| and a2 / |a2|. The lengths of a1 and a2 are equal to the distance d divided by the sine of the angle θ between the two vectors a1 and a2. Therefore, the following equation holds:

[0018]

[0019] According to the definition of the outer product, it is not difficult to obtain:

[0020]

[0021] Where |·| represents the magnitude of the vector, and × represents the cross product of two vectors.

[0022] In this invention, the inner polygon (actually an inner rectangle) is actually a parallel line segment, and the sine of the angle θ between any two adjacent sides should be 1. For two unit vectors a1 / |a1| and a2 / |a2|, the direction of one unit vector can be calculated from the parallel line segment itself, while the other is directly orthogonal to the parallel line segment itself.

[0023] At time t, a certain dilated rectangle in the current image frame can be represented as All the dilated rectangles in the current image frame can be represented as

[0024]

[0025] L t The total number of all parallel line segments in the equation is N.

[0026] Step 3: Extraction of local invariant point features for the dilated rectangle and its tracking in two consecutive frames.

[0027] To utilize feature-based motion estimation, this invention first extracts locally invariant point features from the initial frame, while simultaneously using a region of interest (ROI) formed by fusing all dilated rectangles for range limitation. Therefore, each set of point features detected from any dilated rectangle can be represented as... There are a large number of point features f. t ={x,y}. t ,y t} represents a certain image coordinate in the current frame t. Ψ is the set of all extracted point features in the current image frame. t It can be represented as

[0028]

[0029] Then, the present invention will use these feature points f t ={x,y} is tracked to f in the next frame image. t+1 And any other subsequent frames. In this invention, connecting line segments are drawn on the current frame, connecting all tracked feature points to their initial positions. The set of point features Ψ tracked in the next frame image. t+1 It can be represented as

[0030]

[0031] For any point feature,

[0032] This invention deals with in-vehicle video sequences; therefore, inevitably, the parallel line segments containing the point features have moved in the next frame. In practice, motion estimation based on point features is solved iteratively. Existing motion estimation methods can be performed at different resolutions, thus making the search more efficient and more tolerant of larger displacements. Using the motion estimation method based on point features, the positions of these feature points in the new frame can be accurately obtained.

[0033] Step 4: In the next frame, the tracked local invariant feature points are voted on by each dilated rectangle to achieve matching of salient line segments between two consecutive frames.

[0034] By utilizing the correspondence between identical feature points embedded in two consecutive frames, this invention will extract the salient line segment b from the initial frame. l and b r Matching with all parallel line segments in the next frame enables continuous tracking of salient line segments. Clearly, the new positions of these identical feature points in the next frame should still cluster around the corresponding parallel line segments: feature points detected within an expanding rectangle extending outward from a parallel line segment should, when tracked to the new frame, mostly still fall within the same expanding rectangle of the same parallel line segment. Therefore, this invention achieves the tracking of salient line segment b in two consecutive image frames. l and b r The goal is to automatically track prominent line segments inside highway tunnels over a complete in-vehicle video sequence. This can be achieved by repeating steps 2 through 4 from one frame to another.

[0035] This invention only considers tracking significant line segment b in two consecutive frames. l and b r To achieve significant line segment b l and b r The matching process across two consecutive frames employs a simple voting mechanism. For the i-th feature set tracked from the previous frame... The index i specifically corresponds to two significant line segments b. l or b r The expanded rectangle. Only one number N needs to be accumulated. j This indicates that it belongs to the feature set. Features f t+1 ={x t+1 ,y t+1 It falls into the expanded rectangle at image frame t+1. middle. It is the dilated rectangle of image frame t+1, and the set of dilation matrices Θ from time t+1. t+1 :

[0036]

[0037] In image frame t+1, the total number of parallel line segments is M. Through simple derivation, we can easily obtain:

[0038] k = argmax j N j Clearly, the k-th parallel line segment in the set L of parallel line segments is the one corresponding to the significant line segment b. l and b r The tracking results.

[0039] Compared with the prior art, the present invention has the following significant advantages:

[0040] (1) This invention proposes an automatic tracking method for significant line segments inside highway tunnels based on vehicle-mounted video sequences, providing a new technical means for the efficient acceptance and daily maintenance of tunnel lighting systems.

[0041] (2) The automatic tracking method for significant line segments inside highway tunnels used in this invention is not affected by the jitter of captured video images and still achieves satisfactory results.

[0042] (3) The present invention can effectively improve the performance of detecting the disappearance point in front of the vehicle, thereby greatly improving the availability of existing work. Attached Figure Description

[0043] Figure 1 A general block diagram of the method involved in this invention;

[0044] Figure 2 a-2d Qingdao Jiaozhou Bay Subsea Tunnel LED lighting vehicle-mounted video and complex environment inside the tunnel: (a) transition section; (b) intermediate section; (c) electronic road traffic signs inside the tunnel; (d) LED strobe light for digital camera capture supplementary lighting;

[0045] Figure 3 Robust detection of salient line segments based on geometric structure analysis of road environment inside tunnels;

[0046] Figure 4 Calculate the expanded rectangle for each parallel line segment in the set L of valid parallel line segments;

[0047] Figure 5 Extraction of local invariant feature points for an inflated rectangle and their tracking in two consecutive frames;

[0048] Figure 6In the next frame, the local invariant feature tracking points are voted on by each expansion rectangle to achieve matching of salient line segments between two consecutive frames;

[0049] Figure 7 Automatic tracking results of significant line segments inside highway tunnels based on vehicle-mounted video sequences;

[0050] Figure 8 a-8b Comparison of two methods: the intersection point calculated from two significant line segments and the front vehicle hidden point extracted by the patent: (a) the change of the X coordinate calculated by the two methods with the increase of the number of frames; (b) the change of the Y coordinate calculated by the two methods with the increase of the number of frames. Detailed Implementation

[0051] The specific content of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0052] This invention embodiment was implemented on an Ubuntu 20.04 PC with Qt 5.13 and OpenCV 4.5 installed. The overall block diagram is as follows. Figure 1 As shown. The video footage of the tunnel LED lighting fixture used in this embodiment of the invention was taken in July 2017 by a photographer using a smartphone while traveling through the Qingdao Jiaozhou Bay Subsea Tunnel. Figure 2 As shown, this video records the entire driving process of the main vehicle inside the tunnel, allowing observation not only of the road conditions far ahead of the vehicle but also the LED lighting fixtures arranged on the left and right sides of the tunnel ceiling. The video has a frame rate of 25 frames per second, an image resolution of 720×404, and a total duration of 8 minutes and 48 seconds.

[0053] The Qingdao Jiaozhou Bay Subsea Tunnel is designed as a two-way, six-lane tunnel with a daily traffic volume of up to 53,000 vehicles (2016 statistics). The tunnel is 7,800 meters long, and its internal lighting is divided into several sections, including an enhanced entrance section, the entrance section, and a transition section (see...). Figure 2 a) and the middle section (see Figure 2 b). The internal environment of highway tunnels is relatively complex, including dynamic variable message signs (for issuing warning information), electronic signs, and other brightly lit monitoring and safety facilities, such as... Figure 2 c and Figure 2 As shown in d. When passing through the tunnel during rush hour, the headlights and taillights of other vehicles traveling nearby can also cause significant impact.

[0054] According to the tunnel interior design and construction process requirements, the entire tunnel interior is coated with dark-colored fire-retardant paint. Simultaneously, to ensure wall reflectivity, the tunnel sidewalls are painted with off-white paint up to a height of 3 meters. A sufficiently large area of ​​the tunnel ceiling is entirely dark (but not entirely black). LED lighting fixtures are arranged according to a predetermined pattern within this dark area of ​​the tunnel ceiling. There is a significant difference in grayscale between the dark area of ​​the tunnel ceiling and the white sidewalls. This invention focuses on two key line segments: the two critical left and right boundary lines between the dark area of ​​the tunnel ceiling and the white sidewalls. From... Figure 2 As can be clearly seen, due to significant shaking during filming, the positions of prominent line segments also change considerably between consecutive video frames.

[0055] Step 1: Robust detection of salient line segments based on geometric structure analysis of the road environment inside the tunnel

[0056] Figure 3 The robust detection process for significant line segments based on road environment geometry analysis is demonstrated. Figure 3 Image 'a' is the original video image from the tunnel vehicle. Figure 3 b represents the set of valid parallel line segments in front of the vehicle after simple filtering (shown by the red line). It can be clearly seen that L is mainly reflected on the boundary lines between the various components of the tunnel interior design. Figure 3 c represents the detection result of the vanishing point v in front of the vehicle, indicated by a red dot with a black background in the middle position. From the set of effective parallel line segments L, two key dividing lines—the salient line segment b—can be accurately extracted before the dark area at the top of the tunnel and the white sidewall area. r and b l ,like Figure 3 As shown by the thick blue line in d.

[0057] Step 2: Calculate the expanded rectangle for each parallel line segment in the set L of valid parallel line segments.

[0058] Figure 4 The process of calculating the expanded rectangle for each parallel line segment is shown. Figure 4 'a' represents a polygon and its expanded polygons that are enlarged outwards at equal distances. Figure 4 b is a magnified view of local details to facilitate the description of the specific process of calculating the dilated polygon. Given any parallel line segment in set L (such as...) Figure 4 As shown in c-4d, extending this line segment outwards easily forms an expanding rectangle with an equidistant outward enlargement effect. Clearly, a line segment is a simplified rectangle. The inner rectangle is generated by the prominent line segment, while the outer rectangle is the expanded rectangle. Figure 4 e and Figure 4f displays all the expanded rectangles extending from each parallel line segment. When calculating the expanded rectangle for each parallel line segment in the set L, this invention selects a distance value d of 5 pixels for equidistant expansion. Therefore, each expanded rectangle has a width of 10 pixels and a length equal to the length of the corresponding parallel line segment itself.

[0059] Step 3: Extraction of local invariant point features for the dilated rectangle and its tracking in two consecutive frames.

[0060] Figure 5 The extraction of local invariant feature points facing an inflated rectangle and their tracking process in two consecutive frames are shown. Figure 5 a-5b represents the Region of Interest (ROI) formed by the aggregation of various dilation matrices in two consecutive frames. The first frame is the initial frame, and the second frame is the next frame. Figure 5 c represents the detection result of feature points in the initial frame. The maximum number of locally invariant feature points that can be extracted in each dilated rectangle is 200. When tracking these feature points along the vehicle video sequence, it is inevitable that some of these points will be lost. This will result in the number of tracked feature points decreasing over time. Therefore, it is undoubtedly a good idea to detect new locally invariant feature points based on the remaining number of trackable feature points. Figure 5 c-5d also clearly shows the correspondence between the same feature points in two consecutive frames, represented by the blue line segments on each feature point (pointing from the position of the feature point in the previous frame to the corresponding position in the current frame).

[0061] Step 4: In the next frame, the local invariant feature tracking points vote for each dilated rectangle to achieve matching of salient line segments between two consecutive frames.

[0062] Figure 6 a-6b represents the results of tracking significant line segments in two consecutive frames. Figure 6 c-6d represents the tracking results of salient line segments in two consecutive frames. The two salient line segments tracked are the two yellow line segments in each frame. Figure 7 The results of the salient line segment tracking algorithm for highway tunnels based on vehicle-mounted video sequences, as shown in this embodiment of the invention, are illustrated. In each image, two yellow line segments are located on the left and right sides of the tunnel ceiling, respectively, and completely coincide with the two key dividing lines between the black dark area and the white sidewalls of the tunnel ceiling. These are the results of this embodiment of the invention detecting salient line segments in the initial image frame and continuously tracking them in subsequent image frames. Figure 7 These images in the imagery have fixed time intervals (e.g., 20 frames), therefore Figure 7 The results actually showed significant line segment tracking over 200 consecutive frames. Although Figure 7The selected vehicle-mounted video clips require manual intervention, but not every frame out of the 200 frames yields optimal results. This embodiment of the invention selects to display the tracking results of significant line segments at fixed time intervals, thus minimizing the bias of manual selection. Figure 7 As can be clearly seen, even though the position of significant line segments changes considerably between consecutive video frames due to large shaking during shooting, the significant line segment tracking algorithm implemented in this embodiment of the invention is not affected by the shaking of the captured video image and still achieves satisfactory results.

[0063] The two prominent line segments that this invention focuses on also intersect at the vanishing point in front of the vehicle. Figure 8 Two methods for obtaining the hidden point were compared. One method uses the intersection of two significant line segments as the hidden point (as done in the embodiments of this invention). The other method, as described in the inventor's patent application "A Robust Detection Method for LED Lighting Fixtures in Highway Tunnels" (patent application number 201910443626.2), uses the MSAC method to robustly estimate the set L of effective parallel line segments converging at the hidden point in front of the main vehicle. It can be clearly seen that the X or Y coordinates of the hidden point calculated by both methods change with the increase of the number of frames, such as... Figure 8 As shown. However, the blanking point calculated in this embodiment of the invention will be more accurate, even if there is obvious continuous jitter in the vehicle video image.

[0064] Finally, it should be noted that the above examples are only used to illustrate the present invention and are not intended to limit the technical solutions described in the present invention. Therefore, although the present invention has been described in detail with reference to the above examples, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention. All technical solutions and improvements that do not depart from the spirit and scope of the invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for automatic tracking of salient line segments inside highway tunnels based on vehicle-mounted video sequences, characterized in that, Includes the following steps: Step 1: Robust detection of salient line segments based on geometric structure analysis of the road environment inside the tunnel; The LSD operator is used to extract a set of line segments H representing the geometric structure of the internal environment of a highway tunnel from the original image of the tunnel LED lighting. Here, s is a line segment in the set H, defined by its two endpoints a1 and a2. s={a1,a2}={(x1,y1),(x2,y2)} Where a1 = (x1, y1) and a2 = (x2, y2) represent the start and end points of s, respectively, and x1, y1, x2, and y2 represent the pixel coordinates. Secondly, the line segment set H is filtered to remove lateral parallel lines from the road space ahead and the rear of other vehicles, extracting the effective parallel line segment set L that converges at the vanishing point in front of the main vehicle. The M-sampling and consensus algorithm MSAC is then used to achieve robust detection of the vanishing point v in front of the main vehicle. Finally, in the image space, a new coordinate system XOY is constructed with the vanishing point v in front of the vehicle as the origin. Based on a series of simple conditions, each line segment in the effective parallel line segment set L is further evaluated and filtered, thereby robustly detecting the dark area and white area at the top of the tunnel. sidewall The two key dividing lines on the left and right sides of the wall (b) l and b r， Step 2: Calculate the expanded rectangle for each parallel line segment in the set L of valid parallel line segments; Given time t, set L t Take any parallel line segment and expand it outwards to form an expanding rectangle with an equidistant outward magnification effect. This expanding rectangle consists of a series of straight lines parallel to the sides of the known significant line segment and spaced apart by a distance d. Let l i and l i+1 It is any two adjacent sides of the inner polygon that intersect at vertex p. i， Draw two sides outward from the inner polygon, parallel to l. i and l i+1 The two edges have a spacing of d and are at vertex q. i Intersect, q i It is equal to vertex p in a parallelogram i The sum of the two adjacent sides a1 and a2: q i =p i +(a1+a2) Here, the directions of a1 and a2 are the directions of the corresponding sides of the inner polygon, represented by the difference between the corresponding vertices of the inner polygon. The unit vectors with the same direction as a1 and a2 are represented as a1 / |a1| and a2 / |a2|. The lengths of a1 and a2 are equal to the distance d divided by the sine of the angle θ between the two vectors a1 and a2. According to the definition of the outer product: Where |·| represents the magnitude of the vector, and × represents the cross product of two vectors. At time t, a certain dilated rectangle in the current image frame is represented as All the dilated rectangles in the current image frame are represented as follows: L t The total number of all parallel line segments in the equation is N. Step 3: Extraction of local invariant point features for the expanding rectangle and its tracking in two consecutive frames; First, local invariant point features are extracted from the initial frame. The region of interest (ROI) formed by fusing all dilated rectangles is used to limit the range. Each set of point features detected from any dilated rectangle is represented as F. t i There are a large number of point features f t ={x,y},{x t ,y t } represents a certain image coordinate in the current frame t, and Ψ represents the set of all extracted point features in the current image frame. t Represented as Ψ t ={F t 1 ,F t 2 ,…,F t i ,,…,F t N } These feature points f t ={x,y} is tracked to f in the next frame image. t+1 And in any subsequent frames, connecting line segments are drawn on the current frame, connecting all tracked feature points to their initial positions, forming the set of point features Ψ tracked in the next frame image. t+1 Represented as For any point feature, Step 4: In the next frame, the local invariant feature tracking points vote for each dilated rectangle to achieve matching of salient line segments between two consecutive frames; By utilizing the correspondence between identical feature points embedded in two consecutive frames, the salient line segment b extracted from the initial frame can be... l and b r Matching with all parallel line segments in the next frame achieves continuous tracking of salient line segments. Steps 2 to 4 are repeated to automatically track salient line segments inside highway tunnels in vehicle video sequences. A voting mechanism is used to achieve significant line segment b. l and b r The matching process across two consecutive frames; for the i-th feature set tracked from the previous frame image. The index i specifically corresponds to two significant line segments b. l or b r The expanded rectangle only needs to be increased by a number N. j , indicating that it belongs to the feature set Features f t+1 ={x t+1 ,y t+1 It falls into the expanded rectangle at image frame t+1. middle, It is the dilated rectangle of image frame t+1, and the set of dilation matrices Θ from time t+1. t+1 : In image frame t+1, the total number of parallel line segments is M, from which we obtain: k=argmax j N j The k-th parallel line segment in the set of parallel line segments L is the salient line segment b. l and b r The tracking results.

Images