| [ |
| { |
| "chunk_id": "2919c513-4e13-4c94-ad64-a146c76a4b75", |
| "text": "A Transferable Pedestrian Motion Prediction Model\nfor Intersections with Different Geometries Nikita Jaipuria* Golnaz Habibi* Jonathan P. How\nDept. of Mechanical Engineering Dept. of Aeronautics and Astronautics Dept. of Aeronautics and Astronautics\nMassachusetts Institute of Technology Massachusetts Institute of Technology Massachusetts Institute of Technology\nCambridge, USA Cambridge, USA Cambridge, USA\nnikitaj@mit.edu golnaz@mit.edu jhow@mit.edu", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 0, |
| "total_chunks": 13, |
| "char_count": 452, |
| "word_count": 53, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "d6b78b23-f8cf-4340-932c-bb217d199dc6", |
| "text": "Abstract—This paper presents a novel framework for accurate A B\npedestrian intent prediction at intersections. Given some prior\nknowledge of the curbside geometry, the presented framework2018 can accurately predict pedestrian trajectories, even in new\nintersections that it has not been trained on. This is achieved by\nmaking use of the contravariant components of trajectories in theJun curbside coordinate system, which ensures that the transformation\n!\"($\",&\") of trajectories across intersections is affine, regardless of the\ncurbside geometry. Our method is based on the Augmented Semi ) ) !(($(, &()25\nNonnegative Sparse Coding (ASNSC) formulation [1] and we use\nthat as a baseline to show improvement in prediction performance *\non real pedestrian datasets collected at two intersections in\nCambridge, with distinctly different curbside and crosswalk Curbside Coordinate Frame\ngeometries. We demonstrate a 7.2% improvement in prediction accuracy in the case of same train and test intersections. + +\nFurthermore, we show a comparable prediction performance of !\"- ($\"- ,&\"- )[cs.LG] )′ TASNSC when trained and tested in different intersections with\n!(- ($(- , &(- ) the baseline, trained and tested on the same intersection. Index Terms—Pedestrian intent prediction, skewed coordinate *′\nsystem, Contravariant components, affine transformation, motion\nprimitives, Gaussian Process, sparse coding INTRODUCTION\nFigure 1: An illustration to show how points PA(xA,yA) on the\nIncreased safety of road travelers and a consequent reduction red trajectory in intersection A and PB(xB,yB) on the purple\nin road accident fatality rate has been the main driver of trajectory in intersection B, under the transformation T , map\nresearch on vehicle ADAS and self-driving cars. Recent to points P′A(x′A,y′A) and P′B(x′B,y′B) in the curbside coordinate\nadvances in computation power and an increase in the amount frame. We show that T is in general an affine transformation.\nof publicly available training datasets provided a boost to the Since pedestrian trajectories in urban intersections are signifapplication of state-of-the-art machine learning approaches in icantly constrained by the curbsides, transforming them into\nthis field. the curbside coordinate frame using an affine transformation,\nscenarios requires interaction with multiple moving agents like approximately on top of each other in the curbside coordinate\ncars, cyclists and pedestrians. Intent recognition of pedestrians frame.", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 1, |
| "total_chunks": 13, |
| "char_count": 2490, |
| "word_count": 360, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "40290144-949b-451e-be94-7a5dedc892b8", |
| "text": "This insight helps in developing a general, transferable\nis more challenging than that of cars (and to some extent, pedestrian trajectory prediction model.\ncyclists) because of the absence of pedestrian \"rules of the\nroad\" like staying within road boundaries, following lanes etc. The problem is further complicated when the vehicle-pedestrian\nin new, unseen intersections, with similar situational context\ninteraction occurs in intersection scenarios where additional\nbut varying curbside and crosswalk geometries.\ncontext such as tightly packed sidewalks and traffic lights also\ninfluence pedestrian trajectory. Furthermore, intent modeling, [1] combine the merits of Markovian-based and clusteringin general, is data-intensive. Therefore, there exists a need based techniques to show significant improvement over statefor a general, transferable prediction algorithm, which when of-the-art clustering methods for pedestrian intent estimation.\ntrained on one intersection, can be used for intent prediction However, their approach fails to incorporate context and is\nbased on motion primitives learned using spatial features (x,y\n*These authors contributed equally position in a local reference frame) specific to the training Most of the previous work on context-based in the curbside coordinate frame. TASNSC achieves 7.2%\npedestrian intent recognition is limited to the identification of improvement in prediction accuracy over ASNSC when trained\nstopping versus crossing intent [2]–[7], as opposed to long and tested on the same intersection. When trained and tested on\nterm trajectory prediction which is the aim of our approach. different intersections, TASNSC shows a comparable prediction\nFurthermore, the use of spatial context features like orthogonal performance with the baseline ASNSC trained and tested on\ndistance to curbside [5]–[7] makes these intent classification the same intersection.\nmodels directly dependent on the specific training intersection\nII. PRELIMINARIES\ngeometry and prevents generalization to new intersections with\nvarying curbside and crosswalk geometries. [8] developed a In this section, we first briefly review the trajectory prediction\nmore generic, context-based, multi-model system for predicting approach of [1] which comprises of the ASNSC algorithm\ncrossing behavior in inner-city situations and zebra crossings. for learning motion primitives and a Gaussian Process (GP)\nHowever, the output of their prediction model is again a based framework for future motion prediction using the learned\ncrossing probability as opposed to predicted future trajectory. dictionary of motion primitives. This is followed by a review\nof covariant versus contravariant components of a vector in [9] forecast long-term behavior of pedestrians by making\na general (i.e. including both orthogonal and skewed) two-use of past observed patterns and semantic segmentation of a\ndimensional coordinate system.bird's eye view of the scene. Such an approach, when applied\nin the real world, on board a self-driving vehicle, would require A. Augmented Semi-Nonnegative Sparse Coding (ASNSC)\naccurate high definition semantic priors/maps of each scene\nGiven a training dataset of n trajectories, where each trajecwhich are expensive to create and maintain. It is also unclear if\ntory ti is a sequence of two-dimensional position measurementstheir prediction model can be generalized across new, unseen\ntaken at a fixed time interval ∆t, ASNSC learns a set ofscenes. [10], [11] follow a similar approach to path prediction\nK dictionary atoms, D = [d1,...,dK], in a discretized world,while also demonstrating the ability to \"transfer knowledge\",\nwhere each dk represents a motion primitive (see Fig. 2(a)).and hence, predict in unseen locations with similar semantic\nelements. However, a prior bird's eye view of the scene is B.", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 2, |
| "total_chunks": 13, |
| "char_count": 3845, |
| "word_count": 550, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "cfde237b-9c66-4a23-8ada-0ef5ab92442e", |
| "text": "Trajectory prediction using the learned dictionary\nneeded for both these approaches as well. Our approach, in As shown in 2(b), D is used to segment the original training\ncontrast, is based on learning from real pedestrian trajectories trajectories into clusters, where each cluster is best explained\ncollected by a vehicle equipped with a 3D Lidar and camera. by one of the learned dictionary atoms. A transition matrix,\nIn contrast to previous approaches, the presented approach T ∈ZK×K is thus created, where T(i, j) denotes the number of\nrequires a prior on the curbside geometry only (i.e. angle made trajectories exhibiting a transition from di to dj. A transition\nby intersecting curbs at the corner point of interest) and can be is, therefore, mathematically represented as a concatenation\ngeneralized to any, unseen intersection with similar semantic of two dictionary atoms {di,dj|T(i, j) > 0}. Each transition\ncues as the one trained on. It should be emphasized, however, is modeled as a two-dimensional GP flow field [12], [13]. In\nthat if additional priors, in the form of high fidelity maps, particular, two independent GPs, (GPx,GPy), called GP motion\nare available, they can be easily incorporated in the presented patterns, are used to learn a mapping from the two-dimensional\napproach. position features to the x and y velocities respectively.", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 3, |
| "total_chunks": 13, |
| "char_count": 1361, |
| "word_count": 216, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "64d35c6a-a7c2-42d0-ae79-28615d665429", |
| "text": "The main contributions of this work are as follows: 1) Introduction of a novel representation of distance to\ncurbside as the contravariant components of trajectories\nin the curbside coordinate frame. This representation\nensures that distance to curbside, as a context feature, is\ndependent on curbside geometry only (angle made by\nintersecting curbs).\n(a) (b)\n2) We show that the transformation of trajectories from\nthe original, local frame to curbside coordinate frame\nFigure 2: (a) Each color represents a single dictionary atom dk is affine. It preserves properties such as collinearity,\ni.e. motion primitive; (b) Segmentation of training trajectories parallelism etc. across intersections while encoding\n(in gray) into clusters, where each cluster is best explained by situational context (see Fig. 1).\nthe dictionary atom of the same color in (a). 3) Transferable ASNSC (TASNSC), as a general,\ncontext-based pedestrian intent prediction model for\naccurate prediction in new, unseen intersections with C. Skewed coordinate systems & covariant versus contravarisimilar semantic cues as those that the model is trained ant components of two-dimensional vectors\non. As shown in Fig. 3(a) and Fig. 3(b), a coordinate system\nOur approach, TASNSC is based on the ASNSC framework. can be either orthogonal (represented by unit vectors⃗i,⃗j) or\nIt encodes situational context and provides a general prediction skewed (represented by unit vectors ⃗e1,⃗e2). Covariant and\nmodel by learning motion primitives and their transition contravariant components of a position vector in an orthogonal of pedestrian trajectories i.e. shape and relative distance with\nrespect to curbside is preserved under this transformation (see\nFig. 4 and Fig. 5). Let us define a coordinate frame with its origin\nj at the intersection corner of interest, and its axes along the two\ni curbsides intersecting at the chosen corner as the \"curbside\n(a) (b) (c) coordinate frame\" (see Fig. 1). Given a point P(x,y) in the original, arbitrarily\nFigure 3: (a) Orthogonal coordinate system; (b) Skewed placed local coordinate frame of an intersection (i.e. XY\ncoordinate system; (c) Calculation of contravariant components frame in intersections A and B in Fig. 1), let us define a\nin a skewed coordinate system using trigonometry transformation T : P →P′ s.t. P′(x′,y′) is in the curbside\ncoordinate frame of the same intersection, where x′,y′ are the\ncontravariant components of P′ in the curbside coordinate\ncoordinate system are the same. A position vector in such a\nframe.\nsystem, therefore, has only one representation i.e.⃗r = x1⃗i+y1⃗j\n(see Fig. 3(a)). However, in a skewed coordinate system, the Lemma 1. T is an affine transformation\ncovariant components (x1,y1) and contravariant components\nProof.", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 4, |
| "total_chunks": 13, |
| "char_count": 2773, |
| "word_count": 424, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "9755dfa2-f7cd-4f21-8012-7d2c9b314d7e", |
| "text": "Given the original, orthogonal, local coordinate system(x1,y1) of a position vector do not align. The same position\nO and an intermediate, helper coordinate system H (also\nvector, in such a system, can be represented using both its\northogonal but with its origin at the intersection corner and its\ncovariant and contravariant components. Using the contravariant\nx-axis parallel to the x-axis of the curbside coordinate frame\ncomponents yields ⃗r = x1⃗e1 +y1⃗e2 (see Fig. 3(b)). Since (⃗e1 ·\nC), if TOH and THC represent the coordinate transformation⃗e2) ̸= 0 in a skewed coordinate system, r2 ̸= (x1)2 +(y1)2 in\nfrom O to H and H to C respectively, then T = TOHTHC.general. As shown in Fig. 3(c), basic trigonometric identities\ncan be used for computing the contravariant components of a x′ x x\n=⇒ = T = TOHTHC (3)position vector in a skewed coordinate system. y′ y y x1 = rsin(α −θ)/sinα (1) Since, TOH is simply a combination of rotation and translation,\ny1 = rsinθ/sinα (2) it is an affine transformation. Let us now assume that the\noriginal point P(x,y) in O maps to P∗(x∗,y∗) in H, such that\nSince our aim is pedestrian intent prediction in urban (x∗)2+(y∗)2 = r2. Note that, by definition, the origin and x-axis\nintersections, where curbside geometry significantly constraints of H overlap with the origin and x-axis of C. From Fig. 3(c),\npedestrian motion, learning motion primitives and their transi- if θ is the angle made by the position vector with the x-axes,\ntion in the curbside coordinate frame X′Y ′, as shown in Fig. 1\nx∗= rcosθ,y∗= rsinθ (4)(instead of an arbitrarily placed local coordinate frame XY,\nas in [1]), can help improve prediction accuracies because of Therefore, from (1), (2) and (4), if α is the angle between the\nthe addition of context. Furthermore, in the following section, intersecting curbsides, P′(x′,y′) can be written as\nwe show that pedestrian trajectories, when represented using\ncontravariant components in the curbside coordinate frame, x′ = (rcosθ sinα −rsinθ cosα)/sinα (5)\nundergo an affine transformation across intersections with =⇒x′ = x∗−y∗/tanα (6)\nvarying curbside geometries. This aids us in developing a y′ = rsinθ/sinα = y∗/sinα (7)\ncontext-aware prediction model that can be generalized to any\nintersections. Note that (6), (7) can be combined and written in matrix form\nIII. ALGORITHM x′ x∗ 1 −1/tanα x∗\n= THC = (8) y′ y∗ 0 1/sinα y∗ As discussed earlier, designing a general, transferable prediction model needs features that are independent of the specific For intersections with orthogonal curbsides and therefore an\ntraining intersection geometry. In this section, we show that any orthogonal curbside coordinate frame C, α = π/2 and THC\npoint on a pedestrian trajectory, when mapped from the original, is the identity matrix. Since, THC linearly maps (x∗,y∗) to\narbitrarily placed, local coordinate frame to the curbside (x′,y′), it is an affine transformation.", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 5, |
| "total_chunks": 13, |
| "char_count": 2925, |
| "word_count": 478, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "8c2be94c-5f10-4f68-9739-4e2a5953259c", |
| "text": "Furthermore, since TOH\ncoordinate frame using its contravariant components, undergoes and THC are both affine transformations, T is also an affine\nan affine transformation. The choice of the curbside coordinate transformation by (3).\nframe as the frame in which trajectories are mapped can be\njustified by the fact that pedestrian trajectories are significantly Since T is affine, all general properties of an affine transform\nconstrained by curbsides in intersection scenarios. Since an hold under T , i.e.\naffine transformation preserves properties like collinearity, • Collinearity is preserved\nratios of distances, parallelism etc., the situational context • Parallel lines remain parallel Figure 4: Original (left) and transformed trajectories in the Figure 5: Original (left) and transformed trajectories in the\ncurbside coordinate frame (right) under the transformation T , curbside coordinate frame (right) under the transformation T ,\nwhen the curbs are orthogonal to each other. Trajectories are when the curbs are skewed. Trajectories are shown in blue and\nshown in blue and shaded gray area denotes the sidewalk. shaded gray area denotes the sidewalk. • Convexity of sets is preserved Algorithm 1 Transferable ASNSC (TASNSC)\n• Ratios of distances are preserved i.e. the midpoint of a 1: Input: (⃗e1,⃗e2),Dtr ▷Dtr is the training set of trajectories\nline segment remains the midpoint of the transformed line 2: Training Phase:\n3: for all ti ∈Dtr do segment\n4: t′i = T (⃗e1,⃗e2,ti)\nAs discussed earlier, since the objective of this paper is 5: D′ ←{t′i} ▷D′ is transformed training dataset\npedestrian intent estimation in urban intersections, which 6: end for\nis highly constrained by curbside geometry, transforming 7: D = ASNSC(D′) ▷D is set of learned dictionary atoms\npedestrian trajectories into the curbside coordinate frame helps 8: Testing Phase:\nare curbside unit vectors in testin representing trajectories in different intersection geometries 9: t′o = T (⃗e′1,⃗e′2,to) ▷(e′1,e′2)\nintersection, to is observed trajectory\nin a general frame. This aids in building a context-aware, 10: t′p = predict(d,t′o)\ngeneral prediction model. 11: tp = T −1(t′p)\nAlgorithm 1 describes TASNSC as a transferable version 12: return tp = (x1,y1) ▷predicted trajectory\nof the ASNSC algorithm. We show that TASNSC accurately predicts trajectories in unseen intersections with similar\nsemantics as those that it learned on. Given the curbside as long as the intersection corner is not crowded by obstructions\ncoordinate vectors (⃗e1,⃗e2) in the training intersection, T is such as trees, it is possible to detect the curbside online as the\nused to map the training trajectories from the local, arbitrary vehicle approached the intersection. Real pedestrian trajectories\nplaced coordinate frame to the curbside coordinate frame are collected in two different intersections (see Fig. 8). The\nusing contravariant components.", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 6, |
| "total_chunks": 13, |
| "char_count": 2921, |
| "word_count": 443, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "8c6fe9f0-5fa3-40df-86f9-492660eb3beb", |
| "text": "Motion primitives are then dataset collected in intersection A, with nearly orthogonal\nlearned in the curbside coordinate frame using ASNSC (line curbsides, consists of 186 training and 32 test trajectories\n7). For trajectory prediction in an unseen intersection, first the while that collected in intersection B, with skewed curbsides,\nobserved trajectory is transformed into the curbside coordinate consists of 114 training and 22 test trajectories. An observation\nframe of the test intersection using T (line 9). Motion history of 2.5 seconds prior to the pedestrian entering the\nprimitives and their transition learned in the curbside coordinate intersection is used to predict 5 seconds ahead in time.\nframe of the training intersection are then used for prediction,\nfollowed by a transformation of the predicted trajectory into B. Experiment details\nthe original, local coordinate frame of the test intersection (line Two experiments were conducted for evaluating the predic-\n11). Algorithm 2 describes the procedure for transformation tion performance of TASNSC. In the first experiment, the\nof pedestrian trajectories under T . Fig. 4 and Fig. 5 show training and test intersections are the same. While in the\nthe transformation of trajectories into the curbside coordinate second experiment, the training and test intersections are\nframe under T for an orthogonal and skewed coordinate system\nrespectively. Algorithm 2 Transformation T IV. RESULTS\n1: Input:(⃗e1,⃗e2,ti) ▷curbside unit vectors, trajectoryA. Dataset description\n2: α ←cos−1(e1.e2)\nWe test our algorithm on real pedestrian data collected by 3: for all Pj(xj,yj) ∈ti do\na Polaris GEM vehicle equipped with three Logitech C920 4: xj′ ←sin(α −θ)/sinα ▷refer Fig. 3(c), 0 ≤θ ≤2π\ncameras and a SICK LMS151 LIDARa [14], [15]. A prior 5: yj′ ←sinθ/sinα\n6: end for\noccupancy grip map of the environment, created using the on- 7: return ti′ = {(xj′,yj′)} ▷transformed trajectory\nboard LIDARs, is used to extract curbside boundaries.", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 7, |
| "total_chunks": 13, |
| "char_count": 1996, |
| "word_count": 305, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "a06b9165-e1bf-4d5e-8767-d3cac3dfb3cd", |
| "text": "TASNSC: scene 2, train A, test A 255 training set\nobserved path\npredicited path\n250 actualcurbsidepathleft\ncurbside right 10 15 20 25 30 35 40 45 50\nx (m) Figure 6: Prediction results in intersection A of ASNSC (left), TASNSC trained on the same intersection A (center) and\nTASNSC trained on a different intersection B (right). Ground truth is shown in dotted blue, observed trajectory in pink &\npredicted trajectory in red. In the first scenario (first row), a pedestrian approaches the intersection corner, is faced with a choice\nbetween two crosswalks and decides to continue moving straight. In the second scenario (second row), another pedestrian\napproaches the intersection and is faced with the same choice as in the first, but in this case, decides to turn left.", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 8, |
| "total_chunks": 13, |
| "char_count": 770, |
| "word_count": 128, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "1f212671-9276-46ab-b32f-f35bea62cfa9", |
| "text": "The prediction performance of TASNSC in both TASNSC is either similar to or better than ASNSC when trained\nthese experiments is compared with ASNSC, which we use as and tested on the same intersection. TASNSC also performs\na baseline. Fig. 6 and Fig. 7 show a qualitative comparison well in the case of different training and test intersections. In\nof prediction performance of TASNSC with ASNSC for both those experiments, adding pedestrian traffic light (shown as\nthe experiments in intersections A and B respectively. As is 'tr' in the table) as an additional context feature in the GP\nclear from the trajectory prediction plots, TASNSC improves based transition models [17], boosts prediction performance\nprediction performance over ASNSC in all scenarios when (at the cost of computation time, which is a limitation of the\ntrained and tested on the same intersection. Furthermore, use of the Gaussian Process for Machine Learning (GPML)\nTASNSC shows comparable prediction performance with the package in MATLAB for learning hyperparameters in this case\nbaseline when trained and tested in different intersections. as opposed to manual tuning in the others). Furthermore,\nTable I provides a quantitative comparison of TASNSC the best prediction performance, in terms of both MHD and\nwith ASNSC using two different metrics. The first metric, classification accuracy is achieved by TASNSC when trained\nclassification accuracy represents the percentage of correct and tested in intersection A. This makes sense as the data\npredictions (see Fig. 9) weighted by their likelihood of collected in A is richer in terms of the number of trajectories\nprediction. Mathematically, if a set of n trajectories is pre- and variety in maneuvers/behaviors, which leads to better\ndicted as {t1,...,tn}, with their likelihood of prediction given prediction performance, in general, when trained in A.\nby {l1,...,ln}, and the correct predictions are identified as\n{ti} ∀i ∈C ⊂{1,...,n}, the classification accuracy is given V. CONCLUSION\nby: The presented approach, TASNSC, is a general, accurate\n∑i∈C li\nClassification accuracy % = ×100%. (9) pedestrian trajectory prediction model for urban intersections.\n∑nk=1 lk This is achieved by applying the ASNSC framework for\nThe second metric, Modified Hausdorff Distance (MHD) [16] learning motion primitives and subsequently, modeling the tranis used to compare predicted trajectories with ground truth. As sition between these learned primitives from the transformed\nis clear from the comparison in Table I, TASNSC significantly trajectories in the curbside coordinate frame. The motion\noutperforms ASNSC in classification accuracy, while MHD of primitives and their transition, thus learned, not only encode Figure 7: Prediction results in intersection B of ASNSC (left), TASNSC trained on the same intersection B (center) and\nTASNSC trained on a different intersection A (right). Again, ground truth is shown in dotted blue, observed trajectory in pink\n& predicted trajectory in red.", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 9, |
| "total_chunks": 13, |
| "char_count": 3017, |
| "word_count": 459, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "fefff710-63e0-4937-a541-77e7c76b9715", |
| "text": "In the first scenario (first row), a pedestrian exits the curbside and starts walking along the left\ncrosswalk. In the second scenario (second row), a pedestrian approaches the intersection corner, from inside of the sidewalk\nand continues walking straight to cross the street on the left.", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 10, |
| "total_chunks": 13, |
| "char_count": 289, |
| "word_count": 46, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "8e76a97d-9fef-4c11-acdc-bacfd4a39922", |
| "text": "A B Figure 9: An illustration to\nshow that correct predictions T2\nare defined as those that incorrect\ncorrect predictions\nare within an angular devia- predictions ! (! > 40°)\ntion of 40 degrees from the\nground truth in blue. actual observed\npath path Algorithm Classification MHD Time Train Test tr\nAccuracy (%) (m) (sec) In In\nFigure 8: An overhead snapshot of intersection A with orthogASNSC 84.39 2.267 0.0625 A A Nonal curbsides (left) and intersection B with skewed curbsides\nTASNSC 90.47 2.031 0.0636 A A N\n(right). The training dataset, shown in blue, consists of TASNSC 79.43 2.557 0.0581 B A N\npedestrian trajectories collected using a 3D LIDAR and camera TASNSC 81.73 2.284 0.8643 B A Y\non-board a Polaris GEM vehicle parked at the intersection ASNSC 76.94 2.506 0.0352 B B N\ncorners. TASNSC 82.79 2.637 0.0357 B B N\nTASNSC 75.92 2.95 0.0387 A B N\nTASNSC 79.51 2.859 0.8938 A B Y situational context in the form of distance to curbside, but Table I: Quantitative performance comparison of TASNSC with\nare also agnostic to the specific training intersection geometry. ASNSC\nSuch motion primitives, can therefore, be used for prediction\nin new, unseen intersections with different curbside geometries\nby transforming the observed pedestrian trajectory into the orthogonal curbsides and the other with skewed curbsides.\ncurbside coordinate frame of the test intersection. We test TASNSC shows 7.2% improvement in classification accuracy\nour algorithm on two different intersections, one with almost over ASNSC when trained and tested on the same intersection. A comparable prediction performance, with the baseline, is [14] J. How, \"Predictive positioning and quality of service\nachieved when trained and tested on different intersections. ridesharing for campus mobility on demand systems,\" in Robotics and\nAutomation (ICRA), 2017 IEEE International Conference on. IEEE,\nAddition of traffic light as an additional context feature in the 2017, pp. 1402–1408.", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 11, |
| "total_chunks": 13, |
| "char_count": 1965, |
| "word_count": 311, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "2358cda2-48e3-4ecd-9703-1014b1226980", |
| "text": "GP based transition models helps boost prediction performance [15] J. How, \"Dynamic arrival rate\nin these experiments. estimation for campus mobility on demand network graphs,\" in Intelligent\nRobots and Systems (IROS), 2016 IEEE/RSJ International Conference\nOur approach is limited by the need for a prior on curbside on. IEEE, 2016, pp. 2285–2292.\ngeometry. While one might argue that curbsides can be detected [16] M.-P. Jain, \"A modified hausdorff distance for\non-line as the vehicle approaches an intersection of interest, object matching,\" in Pattern Recognition, 1994. Vol. 1-Conference A:\nComputer Vision & Image Processing., Proceedings of the 12th IAPR\nobservability can be an issue because of occlusions and/or a International Conference on, vol. 1. IEEE, 1994, pp. 566–568.\nlimited FOV of on-board perception sensors. Therefore, there [17] N. How, \"Casnsc: A context-based approach\nis a need to explore the incorporation of uncertainty in curbside for accurate pedestrian motion prediction at intersections,\" in NIPS\nMachine Learning for Intelligent Transportation Systems Workshop\ngeometry in the prediction model and analyze the robustness (MLITS), 2017.\nof TASNSC to it. Furthermore, interaction among pedestrians\nis not considered in the presented TASNSC framework and\nwill be part of future work. Special thanks to Anthony Colangeli, Justin Miller, and\nMichael Everett for their tremendous help in collecting and\nannotating data. This project is funded by a research grant\nfrom the Ford Motor Company.", |
| "paper_id": "1806.09444", |
| "title": "A Transferable Pedestrian Motion Prediction Model for Intersections with Different Geometries", |
| "authors": [ |
| "Nikita Jaipuria", |
| "Golnaz Habibi", |
| "Jonathan P. How" |
| ], |
| "published_date": "2018-06-25", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1806.09444v1", |
| "chunk_index": 12, |
| "total_chunks": 13, |
| "char_count": 1517, |
| "word_count": 225, |
| "chunking_strategy": "semantic" |
| } |
| ] |
