Dynamic Vision for Perception and Control of Motion - Ernst D. Dickmanns Part 11 potx

9.4 Experimental Results 285 curvature of 7.2·10 í5 [1/m] corresponds to a radius of curvature of about 14 km. This means that for 100 m driven the slope change is ~ 0.4°. 7 ·10 -5 5 3 1 -1 C 0v / m -1 100 300 500 runlength / m Figure 9.22. Precise estimation of vertical curvature with simultaneous pitch estimation on underpass of a bridge across Autobahn A8 (Munich – Salzburg) north of Autobahn crossing Munich-south: Top left: Bridge and bottom of underpass can be recognized; top center: vehicles shortly before underpass, shadow of bridge is highly visible. Top right and bottom left: Cusp after underpass is approached; bottom center: leaving the underpass area. Bottom right: Estimated vertical curvature over distance driven: The peak value corresponds to ~ a half degree change in slope over 100 m. Of course, this information has been collected over a distance driven of several hundred meters; this shows that motion stereo with the 4-D approach in this case cannot be beaten by any kind of multiocular stereo. Note, however, that the stereo base length (a somewhat strange term in this connection) is measured by the odometer with rather high precision. Without the smoothing effects of the EKF (double integration over distance driven) this would not be achievable. The search windows for edge detection are marked in black as parallelepipeds in the snapshots. When lane markings are found, their lateral positions are marked with dark-to-bright and bright-to-dark transitions by three black lines looking like a compressed letter capital H. If no line is found, a black dot marks the predicted position for the center of the line. When these missing edges occur regularly, the lane marking is recognized as a broken line (allowing lane changes). 9.4.2.5 Long-distance Test Run Till the mid-1990s most of the test runs served one or a few specific purposes to demonstrate that these tasks could be done by machine vision in the future. Road types investigated were freeways (German Autobahn and French Autoroute), state roads of all types, and minor roads with and without surface sealing as well as with and without lane markings. Since 1992, first VaMoRs and since 1994 also VaMP 286 9 Recursive Estimation of Road Parameters and Ego State while Cruising continuously did many test runs in public traffic. Both the road with lane markings (if present) and other vehicles of relevance had to be detected and tracked; the lat- ter will be discussed in Chapter 11. After all basic challenges of autonomous driving had been investigated to some degree for single specific tasks, the next step planned was designing a vision based system in which all the separate capabilities were integrated into a unified approach. To improve the solidity of the database on which the design was to be founded, a long-distance test run with careful monitoring of failures, and the reasons for failures, had been planned. Earlier in 1995, CMU had performed a similar a test run from the East to the West Coast of the United States; however, only lateral guidance was done autonomously, while a human driver actuated the longitu- dinal controls. The human was in charge of adjusting speed to curvature and keeping a proper distance from vehicles in front. Our goal was to see how poor or well our system would do in fully autonomously performing a normal task of long- distance driving on high-speed roads, mainly (but not exclusively) on Autobahnen. This also required using the speed ranges typical on German freeways which go up to and beyond 200 km/h. Figure 9.23 shows a section of about 38 minutes of this trip to a European project meeting in Denmark in November 1995 according to [Behringer 1996; Maurer 2000]. The safety driver, always sitting in the driver’s seat, or the operator of the computer and vision system selected and prescribed a desired speed according to regulations by traffic signs or according to their personal interpretation of the situation. The stepwise function in the figure shows this input. De- viations to lower speeds occurred when there were slower vehicles in front and lane changes were not possible. It can be seen that three times the vehicle had to decelerate down to about 60 km/h. At around 7 minutes, the safety driver decided to take over control (see gap in lower part of the figure), while at around 17 min- speed measured travel speed set, km/h Figure 9.23. Speed profile of a section of the long-distance trip (over time in minutes) 9.4 Experimental Results 287 utes the vehicle performed this maneuver fully autonomously (apparently to the satisfaction of the safety driver). The third event at around 25 minutes again had the safety driver intervene. Top speed driven at around 18 minutes was 180 km/h (50 m/s or 2 m per video cycle of 25 Hz). Two things have to be noted here: (1) With a look-ahead range of about 80 m, the perception system can observe each specific section of lane markings up to 36 times before losing sight nearby ( L min ~ 6 m), and (2) stopping distance at 0.8 g (-8 m/s 2 ) deceleration is ~ 150 m (without delay time in reaction); this means that these higher speeds could be driven autonomously only with the human safety driver assuring that the highway was free of vehicles and obstacles for at least ~ 200 m. Figure 9.24 gives some statistical data on accuracy and reliability during this trip. Part (a) (left) shows distances driven autonomously without interruption (on a logarithmic scale in kilometers); the longest of these phases was about 160 km. Almost all of the short sequences ( 5 km) were either due to construction sites (lowest of three rows top left), or could be handled by an automatic reset (top row); only one required a manual reset (at ~ 0.7 km). This figure clearly shows that robustness in perception has to be increased sig- nificantly over this level, which has been achieved with black-and-white images from which only edges had been extracted as features. Region–based features in gray scale and color images as well as textured areas with precisely determinable corners would improve robustness considerably. The computing power in micro- Figure 9.24. Some statistical data of the long-distance test drive with VaMP (Mercedes 500 SEL) from Munich to Odense, Denmark, in November 1995. Total distance driven autonomously was 1678 km (~ 95 % of system in operation). 288 9 Recursive Estimation of Road Parameters and Ego State while Cruising Figure 9.25. Typical lateral offsets for manual human steering control over distance driven processors is available nowadays to tackle this performance improvement. The figure also indicates that an autonomous system should be able to recognize and han- dle construction sites with colored and nonstandard lane markings (or even without any) if the system is to be of practical use. Performance in lane keeping is sufficient for most cases; the bulk of lateral offsets are in the range ± 0.2 m (Figure 9.24b, lower right). Taking into account that normal lane width on a standard Autobahn (3.75 m) is almost twice as large as vehicle width, lateral guidance is more than adequate; with humans driving, deviations tend to be less strictly observed every now and then. At construction sites, however, lane widths of down to 2 m may be encountered; for these situations, the flat tails of the histogram indicate insufficient performance. Usually, in these cases, the speed limit is set as low as 60 km/h; there should be a special routine available for handling these conditions, which is definitely in range with the methods developed. Figure 9.25 shows for comparison a typical lateral deviation curve over run length while a human was driving on a normal stretch of state road [Behringer 1996]. Plus/minus 40 cm lateral deviation is not uncom- mon in relaxed driving; autonomous lateral guidance by machine vision compares favorably with these results. The last two figures in this section show results from sections of high- speed state roads driven autonomously in Denmark on this trip. Lane width varies more frequently than on the Autobahn; widths from 2.7 to 3.5 m have been observed over a distance of about 3 km (Figure 9.26). The variance in width estimation is around 5 cm on sections with constant width. Distance in kilometers Figure 9.26. Varying width of a state road can be distinguished from the variance of width estimation by spatial frequency; standard variation of lane width estimation is about 5 cm lane width b in meters Distance in kilometers The left part of Figure 9.27 gives horizontal curvatures estimated for the same stretch of road as the previous figure. Radii of curvature vary from about 1 km to 250 m. Straight sections with curvature oscillating around zero follow sections with larger (constant?) curvature values that are typically perceived as oscillating 9.4 Experimental Results 289 (as in Figure 9.17 on our test track). The system interprets the transitions as clothoid arcs with linear curvature change. It may well be that the road was pieced together from circular arcs and straight sections with step-like transitions in curvature; the perception process with the clothoid model may insist on seeing clothoids due to the effect of low-pass filtering with smoothing over the look-ahead range (compare upper part of Figure 9.17). k Figure 9.27. Perceived horizontal curvature profile on two sections of a high-speed state road in Denmark while driving autonomously: Radius of curvature comes down to a minimum of ~ 250 m (at km 6.4). Most radii are between 300 and 1000 m (R = 1/ c 0 ). The results in accuracy of road following are as good as if a human were driving (deviations of 20 to 40 cm, see Figure 9.28). The fact that lateral offsets occur to the ‘inner’ side of the curve (compare curvature in Figure 9.27 left with lateral offset in Figure 9.28 for same run length) may be an indication that the underlying road model used here for perception may be wrong (no clothoids); curves seem to be ‘cut,’ as is usual for finite steering rates on roads pieced together from arcs with stepwise changes in curvature. This is the price one has to pay for the stabilizing effect of filtering over space (range) and time simultaneously. Roads with real clothoid elements yield better results in precise road following. km Figure 9.28. Lateral offset on state road driven autonomously; compare to manual driving results in Figure 9.25 and curvature perceived in 9.27. [As a historic remark, it may be of interest that in the time period of horse carts, roads used to be made from exactly these two elements. When high-speed cars driven with a finite steering rate came along, these systematic ‘cuts’ of turns by the trajectories actually driven have been noticed by civil engineers who – as a pro- 290 9 Recursive Estimation of Road Parameters and Ego State while Cruising gressive step in road engineering – introduced the clothoid model (linear curvature change over arc length).] 9.5 High-precision Visual Perception With the capability of perceiving both horizontal and vertical curvatures of roads and lanes together with their widths and the ego- state including pitch angle, it is important to exploit precision achievable to the utmost to obtain good results. Sub- pixel accuracy in edge feature localization on the search path has been used as standard for a long time (see Section 5.2.2). However, with good models for vehicle pitching and yawing, systematic changes extended edge features in image sequences can be perceived more precisely by exploiting knowledge represented in the elements of Jacobian matrices. This is no longer just visual feature extraction as treated in Section 5.2.2 but involves higher level knowledge linked to state variables and shape parameters of objects for handling the aperture problem of edge features; therefore, it is treated here in a special section. 9.5.1 Edge Feature Extraction to Subpixel Accuracy for Tracking In real-time tracking involving moving objects, predictions are made for efficient adjustment of internal representations of the motion process with both models for shape and for motion of objects or subjects. These predictions are made to subpixel accuracy; edge locations can also be determined easily to subpixel accuracy by the methods described in Chapter 5. However, on one hand, these methods are geared to full pixel size; in CRONOS, the center of the search path always lies at the center of a pixel (0.5 in pixel units). On the other hand, there is the aperture problem on an edge. The edge position in the search path can be located to sub-pixel accuracy, but in general, the feature extraction mask will slide along a body-fixed edge in an unknown manner. Without reference to an overall shape and motion model, there is no solution to this problem. The 4-D approach discussed in Chapter 6 pro- vides this information as an integral part of the method. The core of the solution is the linear approximation of feature positions in the image relative to state changes of 3-D objects with visual features on their surfaces in the real world. This rela- tionship is given by concatenated HCTs represented in a scene tree (see Section 2.1.1.6) and by the Jacobian matrices for each object–sensor pair. For precise handling of subpixel accuracy in combination with the aperture problem on edges, one first has to note that perspective mapping of a point on an edge does not yield the complete measurement model. Due to the odd mask sizes of 2 n + 1 pixels normal to the search direction in the method CRONOS, mask locations for edge extraction are always centered at 0.5 pixel. (For efficiency reasons, that is, changing of only a single index, search directions are either horizontal or vertical in most real-time methods). This means that the row or column for feature search is given by the integer part of the pixel address computed (designated as ‘entier( y or z)’ here). Precise predictions of feature locations according to some 9.5 High-precision Visual Perception 291 model have to be projected onto this search line. In Figures 9.29 and 9.30, the two predicted points, P* 1N (upper left) and P* 2N (lower right), define the predicted edge line drawn solid. Depending on the search direction chosen for feature extraction (horizontal h, Figure 9.29 or vertical v, Figure 9.30), the nominal edge positions (index N) taking the measurement process into account are m 1hN and m 2hN (9.29) respectively m 1vN and m 2vN (9.30, textured circles on solid line). The slope of the predicted edge is 21 2 1 ()/( ) N NN N N azzyy  . (9.46) For horizontal search directions, the vertical differences 'z 1hN , ' z2hN to the center of the pixel z iN defining the search path are Figure 9.29. Application-specific handling of aperture problem in connection with edge feature extractor in rows (like UBM1; nominal search path location at center of pixel): Basic grid corresponds to 1 pixel. Both predictions of feature locations and measure- ments are performed to subpixel accuracy; Jacobian elements are used for problem specific interpretation (see text). Horizontal search direction: Offsets in vertical direction are transformed into horizontal shifts exploiting the slopes of both the predicted and the measured edges; slopes are determined from results in two neighboring horizontal search paths. 1 pixel y 2N - y 1N P* 1N (y 1N , z 1N ) Horizontal search path 1 Horizontal search path 2 Measured edge 'z 1hN m 2hN m 1hm m 1hN ' y 2hN 'y 1hN 'z 2hN P* 2N (y 2N , z 2N ) 'z 2hN 'y 1pe a N m 2hm m 2hm - m 1hm = ǻm m ǻm N = m 2hN - m 1hN (z 2N - z 1N ) ǻz NN = [entier(z 2N ) – entier(z 1N )] Predicted edge 'y 2pe a m 292 9 Recursive Estimation of Road Parameters and Ego State while Cruising . 1hN 1N 1N 2hN 2N 2N ǻ ()0.5, ǻ (z ) 0.5; z z entier z z z entier     (9.47) in conjunction with the slope a N they yield the predicted edge positions on the search paths as the predicted measurement values /, 1, 2 ihN iN ihN N myzai ' (9.48) In Figure 9.29 (upper left), it is seen that the feature location of the predicted edge on the search path (defined by the integer part of the predicted pixel) actually is in the neighboring pixel. Note that this procedure eliminates the z-component of the image feature from further consideration in horizontal search and replaces it by a corrective y-term for the edge measured. For vertical search directions, the oppo- site is true. For vertical search directions, the horizontal differences to the center of the pixel defining the search path ( 'y 1vN , 'y 2vN ) are ivN iN iN ǻ = ( ) + 0.5 ;y y entier y (9.49) together with the slope a N , this yields the predicted edge positions on the search path as the predicted measurement values to subpixel accuracy: ivN iN ivN N = ǻ , = 1, 2.mzyai (9.50) For point 1 this yields the predicted position m 1vN in the previous vertical pixel (above), while in the case of point 2 the value m 2vN lies below the nominal pixel (lower right of Figure 9.30). Again this procedure eliminates the y-component of the image feature from further considerations in a vertical search and replaces it by a corrective z term for the edge measured. 9.5.2 Handling the Aperture Problem in Edge Perception Applying horizontal search for precise edge feature localization yields the measurement points m 1hm for point 1 and m 2hm for point 2 (dots filled in black in Figure 9.29). Taking knowledge about the 4-D model and the aperture effect into account, the sum of the squared prediction errors shall be minimized by changing the unknown state variable x S . However, the sliding effect of the feature extraction masks along the edges has to be given credit. To do this, the linear approximation of perspective projection by the Jacobian matrix is exploited. This requires that deviations from the real situation are not too large. The Jacobian matrix (abbreviated here as J), as given in Section 2.1.2, approxi- mates the effects of perspective mapping. It has 2· m rows for m features (y and z components) and n columns for n unknown state variables x SD , D = 1, n. Each image point has two variables y and z for describing the feature position. Let us adopt the convention that all odd indices of the 2· m rows (i y = 2·i – 1, i = 1 to m) of J refer to the y-component (horizontal) of the feature position, and all following even indices ( i z = 2·i) refer to the corresponding z-component (vertical). All these couples of rows multiplied by a change vector for the n state variables to be adjusted, Gx SD , D = 1, n yield the changes Gy and Gz of the image points due to Gx S : (, ). T iS i J xyz DD G G G (9.51) 9.5 High-precision Visual Perception 293 Let us consider adjusted image points (y iA , z iA ) after recursive estimation for locations 1 and 2 which have been generated by the vector products 111 11 1 222 222 ; ; ; . ANyS AN zS ANyS ANzS yyJ x zzJ x yyJ x zzJ x DD DD DD DD G  G G G (9.52) These two points yield a new edge direction a A for the yet unknown adjustments Gx SD . However, this slope is measured on the same search paths given by the integer values of the search row (or column) through measurement values m ihm (or m ivm ). The precise location of the image point for a minimum of the sum of the squared prediction errors depends on the Gx SD , D = 1, n, to be found, and it has thus to be kept adaptable. Analogous to Equation 9.46, one can write for the new slope, taking Equation 9.52 into account, 21 2 1 21 21 2 1 2 1 21 21 -(-) - (-) 1/ . 1/ NN z z S AA A AA NN y y S zNS N yNS zz J J x zz a yy yy J J x Jzx a Jyx DD D DD D DD DD G  ' ' G  ' ' G G İ (9.53) For |İ i | << 1, the following linear approximation is valid for the ratio: . (1 ) /(1 ) (1 ) (1 ) 1 1| | 1 2 1 2 1212 12 İİ İİ İİİİİ Applying this to Equation 9.53 yields a linear approximation in Gx SD : 21 21 mod mod 21 21 1 h.o.t. , with / / . yĮ zĮ AN SĮ NN SĮ NN zĮ NyĮ N ǻJ ǻJ aa įxaaCįx ǻz ǻy C ǻJ ǻz ǻJ ǻy ªº §·     |  «» ¨¸ ©¹ ¬¼  (9.54) Horizontal and vertical search paths will be discussed in separate subsections. 9.5.2.1 Horizontal Search Paths The slope of the edge given by Equation 9.46 can also be expressed by the predicted measurement values m 1hN and m 2hN on the nominal search paths 1 and 2 ( dash–dotted lines in Figure 9.29 at a distance ǻz NN = entier(z 2N ) - entier(z 1N ) from each other); this yields the solid line passing through all four points P* 1N , P* 2N , m 1hN and m 2hN . The new term for the predicted slope then is 21 ǻ ()ǻǻ N NN hN hN NN hN azmm zm  . (9.55) Similarly one obtains for the measured slope a m from the two measured feature locations on the same search paths 21 ǻ ()ǻǻ. mNNhmhmNNh azmm zm  m (9.56) Dividing Equation 9.56 by Equation 9.55 and multiplying the ratio by a N yields 21 21 ǻǻ ( ) ( ). mN N hm NhNhN hmhm aamm am m m m    (9.57) Setting this equal to Equation 9.54 yields the relation 21 mod . hN hN mAN NN S m mm aaa aaC x mh D    G ' (9.58) 294 9 Recursive Estimation of Road Parameters and Ego State while Cruising Dividing by a N and bringing the resulting 1 in the form 21 () hm hm m mm ǻ m onto the left side yields after sorting terms, 2121 mod . hN hN hm hm SĮ hm mmmm C įx ǻm   (9.59) With the prediction errors ǻy ipe on the nominal search paths ǻ  ipe ihm ihN ymm, (9.60) Equation 9.59 can be written 12 mod (-)/ . p epe hm S yy mCx D '' ' G (9.61) This is the iteration equation for a state update taking the aperture problem and knowledge about the object motion into account. The position of the feature in the image corresponding to this innovated state would be (n × n vector product of the corresponding row in the Jacobian matrix and the change in the state vector): ;.   iiyĮ SĮ iizĮ SĮ įy ǻJ įx įz ǻJ įx (9.62) Note, however, that this image point is not needed (except for checking progress in convergence) since the next feature to be extracted depends on the predicted state resulting from the updated state ‘now’ and on single step extrapolation in time. This modified measurement model solves the aperture problem for edge features in horizontal search. This result can be interpreted in view of Figure 9.29: The term on the left-hand side in Equation 9.61 is the difference in the predicted and the measured position along the (forced) nominal horizontal search paths 1 and 2 at the center of the pixel. If both prediction errors are equal, the slope does not change and there is no aperture problem; the Jacobian elements in the y-direction at the z-position can be taken directly for computing Gx SD (Gy i ). If the edge is close to vertical ( ǻm m § 0), Equation 9.61 will blow up; however, in this case, ǻy N is also close to zero, and the aperture problem disappears since the search path is orthogo- nal to the edge. These two cases have to be checked in the code for special treat- ment by the standard procedure without taking aperture effects into account. The term on the right-hand side is the modified Jacobian matrix (Equation 9.54). The terms in the denominator of this equation indicate that for almost vertical edges in horizontal search and for almost horizontal edges in vertical search, this formula- tion should be avoided; this is no disadvantage, however, since in these cases the aperture problem is of no concern. 9.5.2.2 Vertical Search Paths The predicted image points P* 1N and P* 2N in Figure 9.30 define both the expected slope of the edge and the position of the search paths ( vertical dash-dotted lines); the distance of the search paths from each other is ǻ y NN = entier(y 2N ) - entier(y 1N ), four pixels in the case shown. The intersections of the straight line through points P* 1N and P* 2N with the search paths define the predicted measurement values (m1vN and m2vN); in the case given, with the predicted image points in the upper right corner of the pixel labeled with index (1, upper left in the figure) and the lower left corner of the pixel labeled (2, lower right in the figure), m1vN lies in the previ- [...]... the behavioral modes and corresponding parameters is performed, yielding the distance from the crossing when to start the sequence of maneuver elements for turning off The fifth phase then is Start of motion control (steering for turnoff maneuver) with continuing (separate) gaze control and perception; this part is rather involved, and timing has to be carefully tuned for the feed-forward components... available for control decision Control output via a sequence of special processors in VaMoRs requires a few tenths of a second; these time delays cannot be neglected for precise steering 10.2 Theoretical Background Before the integrated performance of the maneuver can be discussed, performance elements for motion control of the vehicle (Section 10.2.1), for gaze control (Section 10.2.2), and for recursive... deviations to zero For the final part of the maneuver, the feed-forward control input may be dropped altogether 10.2 Theoretical Background 311 10.2.2 Gaze Control for Efficient Perception As mentioned in Section 10.1.2, there are at least four phases in gaze control for crossroad detection and turnoff in the scheme developed; they are discussed in detail here 10.2.2.1 Attention-controlled Search for Crossroads... real-time system for computer-generated images (CGI, center top) which generates the sequences of images at video rate to be projected onto a curved screen in front of the cameras (left) Machine vision with the original cluster of Transputers 302 10 Perception of Crossroads now closes the loop by sending the control inputs derived both to gaze control (center bottom, real feedback) and for control of. .. Simulation computer controller Two-axis gaze platform cameras environment gaze control Three-axis angular motion simulator (DBS) parallel interface original hardware to be installed in test vehicle Figure 10.4 Hardware-in-the-loop simulation facility of UniBwM for the development of dynamic vision in autonomous vehicles: The shaded area (lower right) shows the original hardware intended for the test vehicle... intertwined activities of perception, gaze control, and vehicle control The first phase is prepare for crossroad perception according to the mission plan, which is a slight turn of gaze direction toward the side of the expected crossroad In the second phase, visual search for candidate features is performed with the favorite methods for feature extraction bottom-up After sets of promising features have... “hardware-in-the-loop” (HIL) testing for guided missiles with infrared sensors Real vision hardware and gaze control was to be part of this advanced simulation intended to allow easy transfer of integrated hardware to test vehicles afterward (shaded area in lower center and right corner of Figure 10.4) It shows the HILsimulation system developed and used for autonomous visual curve steering at UniBwM Before... space to be crossed on the subject’s road of b/2 for turnoffs to the right and 3b/2 for those to the left The actual offset y (and ) at the location of the cg is assumed to be known as well from the separate estimation process running for the road driven It has to be kept in mind that at larger look-ahead distances (of, say, 100, 60, and 20 m), a crossroad of limited width (say, 7 m) appears as a small,... feed-forward components and store results observed during the turnoff maneuver for future evaluation Figure 10.3 shows a typical maneuver from real-world test results by displaying time histories of the camera heading angle, the vehicle heading, and steering angle as well as the summed angle of camera and vehicle heading The maneuver lasts for 27 seconds (s); the last 16 include both gaze and lateral motion. .. other groups, and since, on the other hand, it is essential to have fully repeatable conditions available for testing complex maneuvers and their elements, initial development of curve steering (CS), as it was called, was done in a vision laboratory Hardware-In-the-Loop (HIL) simulation: This at that time unique installation for the development of autonomous vehicles with the sense of vision, derived . crossroad; therefore, speed will be reduced to make more processing time available per distance traveled and for slowing down to the 298 10 Perception of Crossroads speed allowed for curve driving. “hardware-in-the-loop” (HIL) testing for guided missiles with infrared sensors. Real vision hardware and gaze control was to be part of this advanced simulation intended to allow easy transfer of integrated. Background Before the integrated performance of the maneuver can be discussed, performance elements for motion control of the vehicle (Section 10.2.1), for gaze control (Sec- tion 10.2.2), and for