multi view vehicle detection and tracking in crossroads

42
Multi-view Vehicle Detection and Tracking in Crossroads Presented by Alaa Mohammed Khattab Liwei Liu, Junliang Xing, Haizhou Ai Computer Science and Technology Department, Tsinghua University,Beijing 100084, China Email: [email protected]

Upload: aalaa-khattab

Post on 14-Apr-2017

39 views

Category:

Technology


1 download

TRANSCRIPT

Multi-view Vehicle Detection and Tracking in Crossroads

Multi-view Vehicle Detection and Tracking in CrossroadsPresented by Alaa Mohammed KhattabLiwei Liu, Junliang Xing, Haizhou AiComputer Science and Technology Department,Tsinghua University,Beijing 100084, ChinaEmail: [email protected]

ContentIntroductionMultiple View DetectorsTwo-Stage View SelectionDual-layer Complementary Occlusion HandlingExperimentConclusion23-Oct-162

2

IntroductionDetection and tracking of vehicles in traffic scenes is of fundamental importance for surveillance system.

Applications:Traffic analysis.Intelligent scheduling.Abnormal activity detection. 23-Oct-163

IntroductionDifficulties:Gradual and sudden illumination changes.Vehicle view and type changes.Partial and full vehicle occlusions.23-Oct-164

ContentIntroductionMultiple View DetectorsTwo-Stage View SelectionDual-layer Complementary Occlusion HandlingExperimentConclusion23-Oct-165

Traditional MethodsDetect vehicles based on background subtraction.

Track them using techniques like Kalman Fitler and Spatial-Temporal Markov Radom Field with different observations such as contour and appearance

Problems : Sensitive toforeground noise.Camera adjustment.Raining.SnowingShadow.23-Oct-166

6

Multiple View Detectors23-Oct-167

Multiple View DetectorsThe system uses:multiple view detectors

Multiple view detectors : train detectors that cover typical views like frontal (rear) view and side view.

Advantages:Effective for object detection.Robust to illumination variation.23-Oct-168

ContentIntroductionMultiple View DetectorsTwo-Stage View SelectionDual-layer Complementary Occlusion HandlingExperimentConclusion23-Oct-169

Two-Stage View Selection23-Oct-1610

Two-Stage View Selection23-Oct-1611

MMPFMulti-Modal Particle Filter (MMPF) is proposed to track vehicles in explicit view ( frontal or side view ).

MMPF is employed to integrate the two view detectors to track it and perform first stage view selection.23-Oct-1612

MMPFMMPF maintains two groups of particles for a target, one for frontal view and the other for side view

Each particle is evaluated by a confidence reflecting the likelihood of the target belonging to the corresponding view.

23-Oct-1613

MMPF23-Oct-1614

MMPF23-Oct-1615

MMPFDifference between side view total confidences and frontal view total confidences is bigger than a threshold.Side view is the dominant view.23-Oct-1616

MMPFDifference between frontal view total confidences and side view total confidences is bigger than a threshold.Frontal view is the dominant view.23-Oct-1617

MMPFMMPF fail to detect the dominant view.Second stage view selection will be adopted.23-Oct-1618

MMPF23-Oct-1619

MMPFIn the MMPF framework, the observation models need to give a confidence reflecting the likelihood of the target belonging to the corresponding view.

Post Processing: to accurate the confidence of particle.23-Oct-1620

Two-Stage View Selection23-Oct-1621

ST AnalysisSpatial-temporal analysis (ST analysis) is proposed to track vehicles in inexplicit view ( intermediate views between frontal and side view ).23-Oct-1622

ST AnalysisDuring the spatial-temporal analysis, four different types of energy terms are explored to vote for the correct view.

The four energy terms are:Primary particles.Velocity difference.Historical views.Neighboring targets.23-Oct-1623

ST Analysis23-Oct-1624

ST Analysis23-Oct-1625

ST Analysis23-Oct-1626

ContentIntroductionMultiple View DetectorsTwo-Stage View SelectionDual-layer Complementary Occlusion HandlingExperimentConclusion23-Oct-1627

Dual-layer Complementary Occlusion Handling23-Oct-1628

Dual-layer Complementary Occlusion Handling23-Oct-1629

Cluster based detecatedIs used to solve partial occlusion effectively.

Adopt K-Means to cluster confident particles.

Label the particles in one occlusion cluster.

Determine : the particle belongs to which target ?

It is desirable to use the states of the targets in the occlusion cluster last frame (t-1) as the centers of clustering.23-Oct-1630

Dual-layer Complementary Occlusion Handling23-Oct-1631

Backward retracking procedure23-Oct-1632

ContentIntroductionMultiple View DetectorsTwo-Stage View SelectionDual-layer Complementary Occlusion HandlingExperimentConclusion23-Oct-1633

Experiment Settings23-Oct-1634

Detection Performance Evaluation

23-Oct-1635This method give higher detection rate

Tracking Performance EvaluationMT: number of Mostly Tracked trajectories.

Frmt: number of Fragments trajectories.

FAT: number of False trajectories.23-Oct-1636

Complex background23-Oct-1637

Shadow and occlusion

23-Oct-1638

Pedestrian disturbance

23-Oct-1639

ContentIntroductionMultiple View DetectorsTwo-Stage View SelectionDual-layer Complementary Occlusion HandlingExperimentConclusion23-Oct-1640

ConclusionThe system is a real-time system that can deal with view changes and occlusions effectively.

The two-stage view selection is efficient in fusing multiple detectors

Robust toIllumination variation.Different weather conditions (snowy, sunny and cloudy).partial and full occlusions (dual-layer occlusion handling technique).23-Oct-1641

23-Oct-1642