Traffic Sign Localization and Classification Methods: An Overview

Similar documents
Large-Scale Traffic Sign Recognition based on Local Features and Color Segmentation

Face Image Quality Assessment for Face Selection in Surveillance Video using Convolutional Neural Networks

Learning to Detect Faces. A Large-Scale Application of Machine Learning

Face and Nose Detection in Digital Images using Local Binary Patterns

Object Detection Design challenges

Object Detection. Part1. Presenter: Dae-Yong


Generic Object-Face detection

Traffic Signs Recognition using HP and HOG Descriptors Combined to MLP and SVM Classifiers

THE SPEED-LIMIT SIGN DETECTION AND RECOGNITION SYSTEM

Study of Viola-Jones Real Time Face Detector

Detecting and Reading Text in Natural Scenes

Person Detection in Images using HoG + Gentleboost. Rahul Rajan June 1st July 15th CMU Q Robotics Lab

Visual Detection and Species Classification of Orchid Flowers

INTELLIGENT transportation systems have a significant

RGBD Face Detection with Kinect Sensor. ZhongJie Bi

Viola Jones Face Detection. Shahid Nabi Hiader Raiz Muhammad Murtaz

Window based detectors

Recap Image Classification with Bags of Local Features

Automatic Initialization of the TLD Object Tracker: Milestone Update

AN HARDWARE ALGORITHM FOR REAL TIME IMAGE IDENTIFICATION 1

Object Category Detection: Sliding Windows

Object Category Detection: Sliding Windows

Machine Learning 13. week

REAL-TIME ROAD SIGNS RECOGNITION USING MOBILE GPU

Designing Applications that See Lecture 7: Object Recognition

A robust method for automatic player detection in sport videos

Detection of a Single Hand Shape in the Foreground of Still Images

Neural Network and Deep Learning. Donglin Zeng, Department of Biostatistics, University of North Carolina

Out-of-Plane Rotated Object Detection using Patch Feature based Classifier

Detecting Pedestrians Using Patterns of Motion and Appearance (Viola & Jones) - Aditya Pabbaraju

Skin and Face Detection

Haar Wavelets and Edge Orientation Histograms for On Board Pedestrian Detection

Human-Robot Interaction

Deep Learning with Tensorflow AlexNet

Real-time pedestrian detection with the videos of car camera

Face Detection and Alignment. Prof. Xin Yang HUST

Research on Robust Local Feature Extraction Method for Human Detection

Classifier Case Study: Viola-Jones Face Detector

SURF. Lecture6: SURF and HOG. Integral Image. Feature Evaluation with Integral Image

Deep Convolutional Neural Networks. Nov. 20th, 2015 Bruce Draper

A New Strategy of Pedestrian Detection Based on Pseudo- Wavelet Transform and SVM

Adaptive Feature Extraction with Haar-like Features for Visual Tracking

Visual features detection based on deep neural network in autonomous driving tasks

Image Processing. Image Features

Face detection, validation and tracking. Océane Esposito, Grazina Laurinaviciute, Alexandre Majetniak

Visual object classification by sparse convolutional neural networks

Previously. Window-based models for generic object detection 4/11/2011

Fast Traffic Sign Recognition Using Color Segmentation and Deep Convolutional Networks

CAP 6412 Advanced Computer Vision

Real-time Object Detection CS 229 Course Project

CSE 151 Machine Learning. Instructor: Kamalika Chaudhuri

Introduction. How? Rapid Object Detection using a Boosted Cascade of Simple Features. Features. By Paul Viola & Michael Jones

Additive Manufacturing Defect Detection using Neural Networks

Recent Researches in Automatic Control, Systems Science and Communications

Active learning for visual object recognition

Triangle Method for Fast Face Detection on the Wild

Face/Flesh Detection and Face Recognition

Human Detection. A state-of-the-art survey. Mohammad Dorgham. University of Hamburg

Progress Report of Final Year Project

Pedestrian Detection with Deep Convolutional Neural Network

Additive Manufacturing Defect Detection using Neural Networks. James Ferguson May 16, 2016

Face Tracking in Video

Classification of objects from Video Data (Group 30)

Postprint.

Visuelle Perzeption für Mensch- Maschine Schnittstellen

Design guidelines for embedded real time face detection application

Face tracking. (In the context of Saya, the android secretary) Anton Podolsky and Valery Frolov

A Real-Time License Plate Localization Method Based on Vertical Edge Analysis

GPU-based pedestrian detection for autonomous driving

COMBINING NEURAL NETWORKS FOR SKIN DETECTION

Deep Learning Requirements for Autonomous Vehicles

DEPARTMENT OF INFORMATICS

A Cascade of Feed-Forward Classifiers for Fast Pedestrian Detection

Lecture 2 Notes. Outline. Neural Networks. The Big Idea. Architecture. Instructors: Parth Shah, Riju Pahwa

Bayes Risk. Classifiers for Recognition Reading: Chapter 22 (skip 22.3) Discriminative vs Generative Models. Loss functions in classifiers

Fast Human Detection for Intelligent Monitoring Using Surveillance Visible Sensors

Regionlet Object Detector with Hand-crafted and CNN Feature

Object detection using non-redundant local Binary Patterns

Classifiers for Recognition Reading: Chapter 22 (skip 22.3)

CMU Lecture 18: Deep learning and Vision: Convolutional neural networks. Teacher: Gianni A. Di Caro

Object detection as supervised classification

Color Model Based Real-Time Face Detection with AdaBoost in Color Image

Keras: Handwritten Digit Recognition using MNIST Dataset

Ensemble Methods, Decision Trees

Face detection and recognition. Detection Recognition Sally

2 OVERVIEW OF RELATED WORK

Convolutional Neural Networks. Computer Vision Jia-Bin Huang, Virginia Tech

Rapid Natural Scene Text Segmentation

Learning Visual Semantics: Models, Massive Computation, and Innovative Applications

CS 2750: Machine Learning. Neural Networks. Prof. Adriana Kovashka University of Pittsburgh April 13, 2016

Simplifying ConvNets for Fast Learning

A Convex Set Based Algorithm to Automatically Generate Haar-Like Features

Machine Learning for Signal Processing Detecting faces (& other objects) in images

Computer Vision Group Prof. Daniel Cremers. 8. Boosting and Bagging

Computer Vision Group Prof. Daniel Cremers. 6. Boosting

TRAFFIC SIGN RECOGNITION USING A MULTI-TASK CONVOLUTIONAL NEURAL NETWORK

Histogram of Oriented Gradients for Human Detection

Real-time Multi-face Tracking with Labels based on Convolutional Neural Networks

Face Recognition Pipeline 1

Transcription:

Traffic Sign Localization and Classification Methods: An Overview Ivan Filković University of Zagreb Faculty of Electrical Engineering and Computing Department of Electronics, Microelectronics, Computer and Intelligent Systems ivan.filkovic@fer.hr 24.10.2014.

Content Introduction Motivation Traffic Signs Traffic Sign Recognition Traffic Sign Localization State of the Art Integral Channel Features Detector Traffic Sign Classification State of the Art Convolutional Neural Networks Conclusion

Introduction Motivation Motivation Applications Advance Driver Assistance Systems - ADAS Systems for mapping and assessing the state of traffic infrastructure Autonomous vehicles Project VISTA: within the system for detection, tracking and recognition of traffic signs Existing commercial systems Possibility of improving performance Removing restriction like driving in non-urban areas (e.g. only on motorways) Increase detectable and recognizable subset of traffic signs (supported only speed limit signs)

Introduction Traffic Signs Traffic Signs Simple objects, limited by shape and color Noticeable to humans (intense colors, regular forms and reflective surfaces) Problems Changing lighting conditions Partial occlusion Changing weather conditions Different traffic signs perspective Visually similar subsets Standardization

Introduction Traffic Signs Standardization Vienna Convention on Road Signs and Signals (1968) Minor differences between countries 62 signatory countries (Croatia; most European countries) Problem Inability to develop a unified traffic sign recognition system

Introduction Traffic Sign Recognition Traffic Sign Recognition = Traffic Sign Localization + Traffic Sign Classification

Introduction Traffic Sign Recognition Traffic Sign Localization Are there traffic signs in image? Location?

Introduction Traffic Sign Recognition Traffic Sign Classification Determining traffic sign type?

Traffic Sign Localization Traffic Sign Localization Approaches State of the art is based on machine learning techniques Traffic sign localization methods Feature extraction/selection Learning algorithm Problems Localization inaccuracy False alarms Computationally expensive Real time requirement

Traffic Sign Localization State of the Art [Liu, 2014] et. al. presented traffic sign localization framework Novel ideas: 1. Multi-block normalized locale binary pattern (MN-LBP), Tilted MN-LBP features 2. Coarse-to-fine classifier, Split-flow cascade (SFC) 3. Derivate of AdaBoost, Common-finder AdaBoost (CF.AdaBoost)

Traffic Sign Localization State of the Art MN-LBP and Tilted MN-LBP features Express the different patterns found on different traffic signs Similar to classical LBP and MB-LBP Pixel sum values of eight rectangles around the center are compared with an average value to obtain a binary sequence

Traffic Sign Localization State of the Art Split-flow cascade Coarse-to-fine tree structure classifier (improves detection time) Incorporates features shared among multiple classes of traffic signs Branching nodes have same structure as Viola and Jones cascade

Traffic Sign Localization State of the Art Common-finder AdaBoost Finds shared features among different feature sets (efficient SFC-tree) Boosting algorithm selects good features for each training set Sets are then intersected to find same features Shared features are once again trained

Traffic Sign Localization State of the Art [Mathias, 2013] et. al. applied Integral Channel Features detector Detector builds on HOG features and boosted decision trees Very high speed and state-of-the-art performance

Traffic Sign Localization Integral Channel Features Detector Integral Channel Features Detector Sliding window over multiple scales Main focus is on choice of features Diverse information from multiple features Haar-like features Local sums Histograms Integral Channel Features are computed using linear and non-linear transformations from different image channels Learning algorithm is standard boosting

Traffic Sign Localization Integral Channel Features Detector Integral Channel Features Detector Large number of first-order and second-order candidate features are generated randomly First-order features Sum of rectangular regions over different channels Six HOG orientation bins Magnitude of gradient Three LUV color channels Second-order features Weighted sums of first-order features Boosting variant called soft cascade Simple threshold is used after evaluation of every weak classifier Values are set by pruning algorithm which assumes that examples with small weights can be removed early

Traffic Sign Localization Integral Channel Features Detector Integral Channel Features Detector Weak classifiers Decision trees with depth of two Optimization Precomputed features are cached in the beginning of training

Traffic Sign Classification Traffic Sign Classification Approaches Based on machine learning Generic approach Applicable to various object classification Difficult multi-class classification problem Problems Large number of traffic sign classes Wast amount of good quality examples Real time requirement

Traffic Sign Classification State of the Art [Ciresan, 2011] et. al. Committee of Convolutional Neural Networks (CNNs) and a Multi-layer Perceptron (MLP) Preprocessing Examples are resized to 48 x 48 pixels Histogram equalization CNNs architecture Feed-forward architecture Initial weights: sampled from uniform distribution Sub-sampling layers: max-pooling Seven hidden layers 43 neurons in the output layer (one for each class) Neuron Activation function: scaled hyperbolic tangent Output represent class probability Learning algorithm Stochastic gradient descent MPL trained on pre-calculated features (HOG3)

Traffic Sign Classification Convolutional Neural Networks Convolutional Neural Networks Inspired by simple and complex cells in the primary visual cortex of a brain Differ in training procedure and in implementation of convolutional and sub-sampling layers Convolutional layers Number of maps (M) Size of maps (mx,y ) Kernel sizes (kx,y ) Skipping factors (s x,y ) Sub-sampling layers Max-pooling Average-pooling Output Downscaling the output maps of the last convolutional layer to 1 pixel per map Combining the output of last convolutional layer into feature vector

Traffic Sign Classification Convolutional Neural Networks Convolutional Neural Networks mx,y n = mn 1 x,y kx,y n sx,y n + 1 + 1

Conclusion Conclusion Relevant research in traffic sign localization and classification was presented Development of new methods improved results of traffic sign classification and localization Accuracy and real time performance are not satisfying for ADAS application

References Liu, C. and Chang, F. and Chen, Z. (2014) Rapid Multiclass Traffic Sign Detection in High-Resolution Images Intelligent Transportation Systems, IEEE Transactions on PP99, 1 10. Mathias, M. and Timofte, R. and Benenson, R. and Van Gool, L. (2013) Traffic sign recognition - How far are we from the solution? The International Joint Conference on Neural Networks Ciresan, D. and Meier, U. and Masci, J. and Schmidhuber, J. (2011) A committee of neural networks for traffic sign classification The International Joint Conference on Neural Networks

The End