Robotics 2 AdaBoost for People and Place Detection

Robotics 2AdaBoost for Peopleand Place DetectionGiorgio Grisetti, Cyrill Stachniss,Kai Arras, Wolfram Burgardv.1.0, Kai Arras, Oct 09, including material by Luciano Spinello and Oscar Martinez Mozos

Contents Motivation AdaBoost Boosting Features for People Detection Boosting Features for Place Labeling

Motivation: People DetectionPeople tracking is a key technology for manyproblems and robotics-related fields: Human-Robot-Interaction (HRI) Social Robotics: social learning, learning by imitationand observation Motion planning in populated environments Human activity and intent recognition Abnormal behavior detection Crowd behavior analysis and controlPeople detection is a key component therein.

Motivation: People Detection Where are the people?

Motivation: People Detection Where are the people? Why is it hard?Range data contain littleinformation on peopleHard in clutteredenvironments

Motivation: People Detection Appearance of people in range data changesdrastically with:- Body pose- Distance to sensor- Partial occlusion and self-occlusion 2d data from a SICK laser scanner

Motivation: People Detection Appearance of people in range data:3d data from a Velodyne sensor

Motivation: People Detection

Motivation: People DetectionFreiburg main station

Motivation: People Detection Freiburg Main Station data set: raw data

Motivation: People Detection Freiburg Main Station data set: labeled data

Motivation: People Detection Freiburg Main Station data set: labeled data


Boosting Supervised Learning techniqueuser provides Learning an accurate classifier by combiningmoderately inaccurate “rules of thumb” Inaccurate rules: weak classifiers Accurate rule: strong classifier Most popular algorithm: AdaBoost[Freund et al. 95], [Schapire et al. 99]AdaBoost in Robotics:[Viola et al. 01], [Treptow et al. 04], [Martínez-Mozos et al. 05],[Rottmann et al. 05] , [Monteiro et al. 06] , [Arras et al. 07]

AdaBoost What AdaBoost can do for you:1. It tells you what the best "features" are2. What the best thresholds are, and3. How to combine this to a classifier It's a non-linear classifier Robust to overfitting Very simple to implement Classifier design becomes science, not art

AdaBoost A machine learning ensemble technique(also called committee methods) There are other ensemble technique such asBagging, Voting etc. Combines an ensemble of weak classifiers(weak learners) to create a strong classifier Prerequisite: weak classifier better than chance,error < 0.5

ClassificationLinear vs non-linear classifierLNL

ClassificationFormal problem statement: Produce a function that maps Given a training setlabelsvaluesWhat is overfitting? Fitting a statistical modelwith too many parameters Overfitted modelsexplain training data perfectly BUT: cannot generalize!

ClassificationError typesTrue valueTNPredicted valueof the classifierT'N'True PositiveFalse NegativeFalse PositiveTrue Negative Sensivity (True Positive Rate) = TP / TDetectedNot Detected Specificity (True Negative Rate) = TN / Nmany more measures...

AdaBoost: Weak ClassifierAlternativesDecision stump: Single axis-parallel partitionof spaceDecision tree: Hierarchical partition ofspaceMulti-layer perceptron: General non-linear functionapproximatorsRadial basis function: Non-linear expansions basedon kernels

AdaBoost: Weak ClassifierDecision stump Simple-most type of decision tree Equivalent to linear classifier defined by affine hyperplane Hyperplane is orthogonal to axis with which it intersectsin threshold θ Commonly not used on its own Formally,⎧h(x; j,θ) = ⎨+1⎩ −1x j> θelse€where x is training sample, j is dimension

AdaBoost: AlgorithmGiven the training data1. Initialize weights2. For m = 1,...,Mw i=1 n• Train a weak classifier on weighted training dataminimizing€Σ iw iΙ(y i≠ h m(x i))• Compute error of :€€• Compute voting weight of :€€• Recompute weights:€h m(x)h m(x)h m(x)α m= log( 1− e me m)3. Make predictions using the final€strong classifier€n∑e m= w iΙ(y i≠ h m(x i))i =1i =1w i= w iexp{ α m⋅ Ι(y i≠ h m(x i) }n∑w i

AdaBoost: Strong Classifier Computing the voting weightclassifierα mof a weakα mchance: e m = 0.5€α m= log( 1− e m)e m€€error

AdaBoost: Strong Classifier Training is completed...The weak classifiersvoting weightα 1....Mh 1...M(x)are now fixand their The resulting strong classifier is€€⎛ M ⎞H(x i) = sgn ⎜ ∑ α mh m(x i)⎟⎝ m =1 ⎠Class Result {+1, -1}Put your data here€Weighted majority voting scheme

AdaBoost: Weak Classifier Train a decision stump on weighted data This consists in...w iΙ(y i≠ h m(x i))( j * , θ * i=1) = argmin j,θFinding € an optimum parameter θ*for each dimension j =1…d andthen select the j* for which thecost is minimal.n∑n∑i=1w i

AdaBoost: Weak ClassifierA simple training algorithm for stumps:∀ j = 1...dendSort samples x i in ascending order along dimension j∀ i = 1...nendCompute n cumulative sumsThreshold θ j is at extremum ofSign of extremum gives direction p j of inequality€€w cumGlobal extremum in all d sumsthreshold θ* and dimension j*jw cumjw cum(i) =givesi∑k =1w ky k

AdaBoost: Weak ClassifierTraining algorithm for stumps: Intuition Label y :red: +blue: – Assuming allweights = 1jw cum(i) =i∑k =1w ky kθ * , j * = 1

AdaBoost: AlgorithmGiven the training data1. Initialize weights2. For m = 1,...,Mw i=1 n• Train a weak classifier on weighted training dataminimizing€Σ iw iΙ(y i≠ h m(x i))• Compute error of :€€• Compute voting weight of :€€• Recompute weights:€h m(x)h m(x)h m(x)α m= log( 1− e me m)3. Make predictions using the final€strong classifier€n∑e m= w iΙ(y i≠ h m(x i))i =1i =1w i= w iexp{ α m⋅ Ι(y i≠ h m(x i) }n∑w i

AdaBoost: Step-By-Step Training data

AdaBoost: Step-By-Step Iteration 1, train weak classifier 1Thresholdθ* = 0.37Dimensionj* = 1Weighted errore m = 0.2Voting weightα m = 1.39Total error = 4

AdaBoost: Step-By-Step Iteration 1, recompute weightsThresholdθ* = 0.37Dimensionj* = 1Weighted errore m = 0.2Voting weightα m = 1.39Total error = 4



AdaBoost: Step-By-Step Iteration 3, train weak classifier 3Thresholdθ* = 0.14Dimension, signj* = 2 , negWeighted errore m = 0.25Voting weightα m = 1.11Total error = 1

AdaBoost: Step-By-Step Iteration 3, recompute weightsThresholdθ* = 0.14Dimension, signj* = 2 , negWeighted errore m = 0.25Voting weightα m = 1.11Total error = 1











AdaBoost: Step-By-Step Final Strong ClassifierTotal error = 0!...which is rarelymet in practiceColored backgrounddots infigure is test set.

AdaBoost: Wrap Up Misclassified samples receive higher weights The higher the weight the "more attention"a training sample gets Algorithm generates weak classifier by training thenext learner on the mistakes of the previous one Now we also understand the name:adaptive Boosting AdaBoost Once training is done, AdaBoost is a voting method Large impact on ML community and beyond


Problem and Approach Can we find robust features for people, legs andgroups of people in 2D range data? What are the best features for people detection? Can we find people that do not move?Approach: Classifying groups of adjacent beams Computing a set of simple (scalar) features onthese groups Boosting the features

Related Work People Tracking[Fod et al. 2002][Kleinhagenbrock et al. 2002][Schulz et al. 2003][Scheutz et al. 2004][Topp et al. 2005][Cui et al. 2005][Schulz 2006][Mucientes et al. 2006]SLAM in dyn. env.[Montemerlo et al. 2002][Hähnel et al. 2003][Wang et al. 2003]... People detection done with very simple classifiers:manual feature selection, heuristic thresholds Typically: narrow local-minima blobs that move

Segmentation Divide the scan into segments"Range image segmentation"

SegmentationRaw scanSegmented scan Method: Jump distance conditionSee [Premebida et al. 2005] for survey Size filter:rejection of too small segmentsFeature profiles



FeaturesSegment1. Number of points2. Standard Deviation3. Mean avg. deviation from median4. Jump dist. to preceding segment5. Jump dist. to succeeding segment6. Width

FeaturesSegmentr c7. Linearity8. Circularity9. Radius

FeaturesSegment10. Boundary Length11. Boundary Regularity12. Mean curvature13. Mean angular difference14. Mean speed

Features Resulting feature signature for each segment

Training: Data Labeling Mark segments that correspond to people Either manuallyor automatically

Training: Data Labeling Automatic labeling: obvious approach, define areaof interest3 m4 m Here: discrimination from clutter is interesting.Features include spatial relation between fore- andbackground. Thus: labeling is done by hand

Training Resulting Training Set+1 -1Segments correspondingto people(foreground segments)Segments correspondingto other objects(background segments)

AdaBoost: Weak Classifiers Each binary weak classifier h j (x) is created using asingle-valued feature f j in the form: Where θ j is a threshold and p j ∈ {-1, 1} indicates thedirection of the inequality Weak classifier must be better than random

AdaBoost: Final Strong ClassifierStrong Binary Classifierexample 1 . . . example Nw 1 h 1Boosting...Σ{-1,1}w T h Tf #1 . . . f #14Vocabulary of featuresWeighted majorityvote classifier

ExperimentsEnv. 1: Corridor, no clutterEnv. 2: Office, very cluttered

ExperimentsEnv. 1 & 2: Corridor and OfficeEnv. 1→2: Cross-evaluationTrained in corridor, applied in office

ExperimentsAdding motion feature (mean speed, f#14)→ Motion feature has no contributionExperimental setup: Robot Herbert SICK LMS 200 laser rangefinder, 1 deg angularresolution

Experiments Comparison with hand-tuned classifier Jump distance θ δ = 30 cm Width θ w,m = 5 cm, θ w,M = 50 cm Number of points θ n = 4 Standard deviation θ σ = 50 cm Motion of points θ v = 2 cmPeople are often not detected

ExperimentsFive best features:1: Radiusof LSQ-fitted circle, robust size measure (#9)2: Mean angular differenceConvexity measure (#13)3/4: Jump distancesLocal minima measure (#4 and #5)5: Mad from medianRobust compactness measure (#3)

Result: ClassificationTFT F

Summary People detection phrased as a classification problem ofgroups of neighboring beams AdaBoost allows for a systematic approach toperform this task One-shot/single-frame people detection with over90% accuracy Learned classifier clear superior to hand-tunedclassifier No background knowledge such as an a priori map isneeded (e.g. to perform background substraction)


Place Labeling: Motivation A map is a metric and topological modelof the environment

Place Labeling: Motivation Wanted: semantic information about placesRoomCorridorDoorway

Scenario ExampleAlbert, whereare you?I am in thecorridor

Scenario Example 2 Semantic mappingCorridorRoomDoorway Human-Robot Interaction of type:"Robot, get out of my room, go into the corridor!"

Problem Statement Classification of the position of the robot usinga single observation: a 360° laser range scan

Observations

ObservationsRoom

ObservationsRoom

ObservationsRoomDoorway

ObservationsRoomDoorway

ObservationsCorridorRoomDoorway

Similar Observations

Similar ObservationsCorridorDoorway

Classification Problem

Classification Problem

Classification Problem?CorridorRoomDoorway

Representing the Observations How we represent the 360 laser beams for ourclassification task? As a list of beamsProblem: which beam is the first beam?Not invariant to rotation!!=

Representing the Observations A list of scalar geometrical features of the scanThe features are all invariant to rotation=

Simple Featuresd idd•Σ d i• Gap = d > θ• f = # GapsMinimum• f = dd• f =Area • f =Perimeter • f = d

Simple Features Features of the raw beams

Simple Features Features of the closed polynom P(z)made up by the beams

Multiple ClassesCorridorRoomDoorway1 23

Multiple ClassesCorridorRoomDoorway1 23

Multiple Classes Sequence of binary classifiers in a decision listCorridorClassifierH(x)=–1RoomClassifierH(x)=–1DoorwayH(x)=1CorridorH(x)=1Room Order matters as accuracy differs Order according to error rate Generalizes to sequential AdaBoost for K classes

ExperimentsTraining (top)# examples:16045Test (bottom)# examples:18726classification:93.94%Building 079Uni. FreiburgCorridorRoomDoorway

ExperimentsTraining (left)# examples:13906Test (right)# examples:10445classification:89.52%Building 101Uni. FreiburgCorridorRoomDoorwayHallway

Application to New EnvironmentTraining mapIntel Research Labin Seattle

Application to New EnvironmentTraining mapIntel Research Labin SeattleCorridorRoomDoorway

Training Learn a strong classifier from a set of previouslylabeled observationsRoomCorridorDoorwayfeaturesfeaturesfeaturesAdaBoostMulti-class Classifier

Robotics 2 AdaBoost for People and Place Detection

Create successful ePaper yourself

Delete template?

Save as template?