Aucun titre de diapositive

Alerting the Drivers about Road Signs with Poor Visual Saliency. Ludovic SIMON ... Road Sign Appearance Learning: Uses Support Vector Machine (SVM).
694KB taille 2 téléchargements 72 vues
Alerting the Drivers about Road Signs with Poor Visual Saliency

Poster Session: Vision Systems, Driver Modeling and Monitoring - Poster I Topics: Driver Assistance Systems Paper ID: 872404

Ludovic SIMON, Jean-Philippe TAREL, Roland BREMOND Laboratory for Road Operation, Perception, Simulation and Simulators Université Paris EST, LEPSIS, INRETS-LCPC [email protected], [email protected], [email protected]

CONTEXT Solution : An ADAS

Road Sign Saliency Evaluation : A need ●

Saliency: Degree to which an object attracts visual attention for a given background.



Prevents drivers from missing road signs necessary for road safety.



Road signs: act on road safety and traffic control (driver's guidance, alert, information).



Use a computational saliency estimator (SCS) of the road sign saliency.



Accident Risk: Not all of them are seen by all drivers.



Alert only in case of poor road sign saliency.

Saliency Problem.

Road Images

World Onboard Camera

Road Sign Saliency

Road Sign Image Road Signs Detector

Saliency Estimator

Alert On Poorly Salient Road Signs

Sign to Display Risk Decision

HUD

Size of the Road Sign

STATE OF THE ART

Saliency Estimator ●

Road Signs Detector

Several constraints ●

Good detection rate: No signs missed.



Low False Detection Rate.









Robust: partial occlusions, shadows, lighting changes, perspective distortion and low visibility.



Must compute the size of the detected road sign.

Must be correlated to Driver's Visual Attention Computational Model of visual attention (Itti, 1998) : ➔ Works for a Bottom-Up Task (image memorization) ➔ Not Valid for a Top-down Task (search for a sign) Model of Visual Search : ➔ mainly theoretical rather than computational (Navalpakkam, 2005; Sundsted, 2005) ➔ computational one limited to laboratory situation (Gao, 2004; Le Meur, 2006; Wolfe, 2007)

A SVM-Based Saliency Model Road Sign Appearance Learning: Uses Support Vector Machine (SVM). ● Training Stage: Compute a classification function C(x) by learning positives samples (road signs) and negatives ones. ●

Proposed Approach

● ●

Itti's Saliency Map

One important Driving Sub-Task : looking for road sign (ie search task)

PROPOSAL



Constraints and Problems

Capture the priors a driver learns about the appearance of any (set of) road signs. Relying on statistical learning algorithm.

● ●

Define the saliency of the road sign by the confidence of the algorithm results

Testing Stage: C(x) applied to image windows give by the road sign detector Intrinsical Computed Saliency Map (ICS): ➔ Maximum confidences values C(x) over various sliding windows sizes. ➔ Subtraction of the mean over a background-windows, 2° of visual angle around each detected sign

Map of C(x) with Scale 40 x 40 ●

Proposed Visual Search Saliency Map

Map of C(x) with Scale 10 x 10

Size-dependent Computed Saliency (SCS): ➔ SCS(i): Visual Saliency of the road sign i ➔

ICS(i): Intrinsical Saliency of the road sign i



A(i): Area of the road sign i

Resulting ICS Map

4

SCS  i=  ICS  i× A i

VALIDATION Reference Data ●



Psycho-Visual experiment: ➔ Remote Eye Tracker. ➔ 40 images, 76 « no entry » signs ➔ various saliency levels. Objective Evaluation of the saliency: ➔ subject's eye fixation. ➔ D i , j  : 1 if sign i is noticed by subject j. ➔



Statistical Analysis

Typical scan-path of one subject searching for « no entry » sign.

HDR i =∑ jj =32 =1 D i , j/ 32

Subjective Evaluation of the saliency: subjects' scoring. ➔ score i , j : rate for subject j and sign i. ➔



SSS i , j=

score  i , j − E i  score  i , j   i  score i , j 

5

Human Detection Rate (HDR) as a function of the Subject Standardized Score (SSS).

Size-dependent Computed Saliency (SCS) as a function of the Subject Standardized Score (SSS).

HDR and SSS are linked. Signs which are not well notices are above a score threshold of 4. ● There is a linear relation between SSS and SCS, our computational model. ● Thus, our model of visual search saliency of road signs is correlated to Human detection Rate. ● The proposed model (SCS) explains 56% of the variance between signs. ●