Two-Stage Pedestrian Detection Using HOG and FMT Descriptors

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for detecting road objects like pedestrians in vehicle images face challenges due to variability in human appearance, poses, and illumination, leading to high false detection rates, especially in the second validation stage.

Innovation Solution

A two-stage method for road object recognition using a first stage of detection with Adaboost algorithm and Haar features, followed by a second stage of validation combining Histogram of Oriented Gradients (HOG) and Fourier-Mellin Transformation (FMT) descriptors to classify regions into pedestrians or non-pedestrians, with a support vector machine classifier to reduce false alarms.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If a fast search mechanism based on simple features is applied in the first stage of detection, then detection speed is improved, but the number of false detections increases

Engineering Contradiction:
Improvedetection speedVSAvoidfalse detection rate
Core Design Contradiction:
SpeedVSReliability

Solution Approach 1:

The detection process is divided into two distinct stages: a first stage using simple features for fast detection, and a second stage using complex features for validation. This segmentation allows the system to achieve high speed in the first stage while reducing false detections through the more rigorous second stage, resolving the contradiction between speed and reliability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The first stage performs preliminary detection using simple features to quickly identify potential pedestrian regions before applying the more computationally intensive validation in the second stage. This preliminary action filters out obvious non-pedestrian regions early, maintaining speed while setting the stage for more accurate classification.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If a set of features based on histograms of oriented gradients is used in the second stage of validation, then classification accuracy is improved, but the detection speed decreases

Engineering Contradiction:
Improveclassification accuracyVSAvoiddetection speed
Core Design Contradiction:
Measurement precisionVSSpeed

Solution Approach 1:

The feature extraction process is segmented into two stages with different computational complexities. The first stage uses simple features for rapid processing, while the second stage applies HOG features only to regions of interest identified in the first stage. This segmentation maintains high classification accuracy where needed while preserving overall detection speed.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The computationally intensive HOG feature extraction is applied partially only to candidate regions identified in the first stage, rather than to the entire image. This partial action achieves high classification accuracy for potential pedestrians while avoiding the speed penalty of processing the complete image with complex features.

Inventive Principle:
Principle #16Partial or excessive action

3Reliability

If a classifier is applied to verify hypotheses in the second stage of validation, then false detections are reduced, but the processing time increases

Engineering Contradiction:
Improvefalse detection reductionVSAvoidprocessing time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The verification process is segmented into two stages: a fast first stage that quickly identifies candidate regions, and a second stage that applies classification only to these candidates. This segmentation reduces the overall processing time by avoiding full classification of the entire image while still achieving reliable false detection reduction through the second stage classifier.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The first stage performs preliminary filtering to identify candidate pedestrian regions before the second stage classifier is applied. This preliminary action reduces the number of regions requiring full classification, thereby reducing processing time while maintaining the reliability benefits of the classifier for the regions that do undergo full verification.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP2583216B1Method of predetermined road object recognition in images and vehicle driving assistance device
Publication Date: 2017.09.13 VALEO SCHALTER & SENSOREN GMBH
  • EP2583216B1 patent drawingFigure 1
  • EP2583216B1 patent drawingFigure 2~5
  • EP2583216B1 patent drawing

AI summary

The invention relates to a method of predetermined road object recognition in images, including : - a first stage of detection (101), to detect predetermined road object-like regions in images, done by sliding a search window through an image and checking whether an image region is classified as said predetermined road object or non-predetermined road object, and - a second stage of validation (102), to classify said predetermined road object-like regions into predetermined road objects and non- predetermined road objects, characterized in that in said second stage of validation (102), first descriptors computed inside said predetermined road object-like region from extracting features based on histogram of oriented gradients (HOG), and second descriptors from a Fourier-Mellin transformation (FMT) on the predetermined road object-like region pixels, supplies both a classifier to classify said predetermined road object-like regions into predetermined road objects and non-predetermined road objects. The invention also relates to a vehicle driving assistance device to implement said method of predetermined road object recognition.