CN103093214B - A kind of pedestrian detection method based on vehicle mounted infrared camera - Google Patents

Abstract

The present invention relates to a kind of pedestrian detection method based on vehicle mounted infrared camera.Traditional Measures compare is consuming time.First the present invention sets up infrared camera in automobile head cursor position level and carries out camera calibration.Secondly preset camera calibration distance range, determine up-and-down boundary.Then screen candidate's bounding box, extract the HOG feature of residue candidate bounding box, and utilize the SVM classifier of pre-training to determine whether pedestrian.Finally utilize the bounding box being judged to be pedestrian in yardstick and distance sorting procedure 5, after level and smooth, obtain final pedestrian's detection and tracking.The present invention is consuming time low, and false drop rate is low, is convenient to embedded real-time application.

Description

A kind of pedestrian detection method based on vehicle mounted infrared camera

Technical field

The invention belongs to field of intelligent video surveillance, be specifically related to a kind of pedestrian detection method based on vehicle mounted infrared camera.

Background technology

For pedestrian detection under in-vehicle camera, current method or the background subtraction [1] based on moving camera, or based on the mode identification method [2] that full figure is searched for, be all method more consuming time.Based on the method [3] of two (many) orders camera although detect after extraction of depth information area-of-interest can be utilized, obtain a depth information inherently process consuming time.

Some methods [4] by extracting some feature of people, as head etc. carry out Primary Location after carry out again detecting reducing consuming time, but degradation problem under often causing verification and measurement ratio.

Pertinent literature and patent:

[1]Online Moving Camera Background Subtraction. Ali Elqursh, Ahmed Elgammal, ECCV, 2012.

[2] based on pedestrian detection method and pedestrian's anti-collision early warning system of machine vision, publication number: 102765365A.

[3]Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite. Andreas Geiger, Philip Lenz, Raquel Urtasun, CVPR, 2012.

[4]Pedestrian Detection and Tracking in Far Infrared Images. Masahiro Yasuno, Noboru Yasuda,Masayoshi Aoki, ITSC, 2005.

Summary of the invention

For the problems referred to above, the present invention uses infrared camera and introduces dynamic programming algorithm, adopts the method detected after first extracting area-of-interest, meets consuming time and requirement that is accuracy simultaneously.

The technical scheme that technical solution problem of the present invention is taked is:

1., at automobile head cursor position level erection infrared camera, its height value is 0.5-0.7 rice.

2. calibration for cameras, specifically: measure the actual antenna height H of camera, namely camera optical axis is to the distance on ground.Find after spacious level land stopped vehicle, call calibrating procedure mark local horizon and obtain camera calibration result after inputting camera antenna height.The ratio claiming image ordinate and picture altitude is image normalization ordinate, and calibrating procedure need input camera heights H and horizontal image normalization ordinate Y0, camera vertical field of view angle , preset pedestrian average height DH=1.7m.Coefficient is:

，

Wherein .

Suppose that the horizontal range that on ground, certain puts camera is Z1, then the image normalization ordinate of its correspondence .Similarly, suppose that certain point on height DH surface level is Z2 to the horizontal range of camera, then the image normalization ordinate of its correspondence .

3. preset camera calibration distance range, the longitudinal coordinate Y obtained by step 2 and the formula of horizontal range Z, the image normalization ordinate scope corresponding to detection level distance range preset and object height to be detected corresponding to image ordinate can be obtained.Within the scope of ordinate, candidate's bounding box is generated, by all candidate's bounding boxs up and down by preset displacement translation 3-4 group according to default horizontal and vertical step-length.

4. after every two field picture input, first by DP algorithm determination up-and-down boundary.Carry out from left to right within the scope of the ordinate obtained in step 3 during planning, cost function adopts the gradient in band direction, column direction allows step to a certain degree.

5. obtain up-and-down boundary and also just determine area-of-interest place, according to the candidate's bounding box obtained in this region screening step 3, extract the HOG feature of residue candidate bounding box, and utilize the SVM classifier of pre-training to determine whether pedestrian.

6. utilize the bounding box being judged to be pedestrian in yardstick and distance sorting procedure 5.

7. after the result of step 6 is level and smooth in time, obtain final pedestrian's detection and tracking effect.

8. do corresponding display and warning according to step 7 result.

Beneficial effect of the present invention: the present invention is consuming time low, and false drop rate is low, is convenient to embedded real-time application.

Accompanying drawing explanation

Fig. 1 is the inventive method process flow diagram.

Embodiment

Below in conjunction with drawings and Examples, the invention will be further described.

As shown in Figure 1, the inventive method comprises the following steps:

1., at automobile head cursor position level erection infrared camera, it is highly according to concrete vehicle slightly difference, generally at 0.5-0.7 rice.

2. calibration for cameras, specifically: measure the actual antenna height H of camera, namely camera optical axis is to the distance on ground.Find after spacious level land stopped vehicle, call calibrating procedure mark local horizon and obtain camera calibration result after inputting camera antenna height.The ratio claiming image ordinate and picture altitude is image normalization ordinate, and calibrating procedure need input camera heights H and horizontal image normalization ordinate Y0, camera vertical field of view angle , preset pedestrian average height DH=1.7m.

Coefficient is ,

Wherein .

Suppose that the horizontal range that on ground, certain puts camera is Z1, then the image normalization ordinate of its correspondence .Similarly, suppose that certain point on height DH surface level is Z2 to the horizontal range of camera, then the image normalization ordinate of its correspondence .

3. preset camera calibration distance range, the longitudinal coordinate Y obtained by step 2 and the formula of horizontal range Z, the image normalization ordinate scope corresponding to detection level distance range preset and object height to be detected corresponding to image ordinate can be obtained.Within the scope of ordinate, generate candidate's bounding box according to default horizontal and vertical step-length, consider that pedestrian level fluctuation and vehicle travel process the caused image that jolts fluctuate up and down, by all candidate's bounding boxs up and down by preset displacement translation 3-4 group simultaneously.

4. after every two field picture input, first by DP algorithm determination up-and-down boundary.Carry out from left to right within the scope of the ordinate obtained in step 3 during planning, cost function adopts the gradient in band direction, column direction allows step to a certain degree to allow some ladders on human body, as head shoulder etc.

5. obtain up-and-down boundary and also just determine area-of-interest place, according to the candidate's bounding box obtained in this region screening step 3, extract the HOG feature of residue candidate bounding box, and utilize the SVM classifier of pre-training to determine whether pedestrian.

6. utilize the bounding box being judged to be pedestrian in yardstick and distance sorting procedure 5.

7. after the result of step 6 is level and smooth in time, obtain final pedestrian's detection and tracking effect.

8. do corresponding display and alarm lamp according to step 7 result.

The above; be only preferred embodiment of the present invention, be not intended to limit protection scope of the present invention, should understanding be with; the present invention is not limited to implementation as described herein, and the object that these implementations describe is to help those of skill in the art to put into practice the present invention.

Claims (1)

1., based on a pedestrian detection method for vehicle mounted infrared camera, it is characterized in that the method comprises the following steps:

Step 1. is at automobile head cursor position level erection infrared camera, and its height value is 0.5-0.7 rice;

Step 2. calibration for cameras, specifically: measure the actual antenna height H of camera, namely camera optical axis is to the distance on ground; Find after spacious level land stopped vehicle, call calibrating procedure mark local horizon and obtain camera calibration result after inputting camera antenna height; The ratio claiming image ordinate and picture altitude is image normalization ordinate, and calibrating procedure need input camera heights H and horizontal image normalization ordinate Y0, camera vertical field of view angle α, presets pedestrian average height DH=1.7m; Coefficient is:

$\mathrm{Coef}0=Y0,\mathrm{Coef}1=H\frac{{t2}^2+1}{2t1},\mathrm{Coef}2=(\mathrm{DH}-H)\frac{{t2}^2+1}{2t1}$

Wherein suppose that the horizontal range that on ground, certain puts camera is Z1, then the image normalization ordinate of its correspondence similarly, suppose that certain point on height DH surface level is Z2 to the horizontal range of camera, then the image normalization ordinate of its correspondence $Y2=\mathrm{Coef}0-\frac{\mathrm{Coef}2}{Z2};$

Step 3. presets camera calibration distance range, the formula of longitudinal coordinate Y0, Y1, Y2 of being obtained by step 2 and horizontal range Z1, Z2, can obtain the image normalization ordinate scope corresponding to detection level distance range preset and object height to be detected corresponding to image ordinate; Within the scope of ordinate, candidate's bounding box is generated, by all candidate's bounding boxs up and down by preset displacement translation 3-4 group according to default horizontal and vertical step-length;

After the every two field picture input of step 4., first by DP algorithm determination up-and-down boundary; Carry out from left to right within the scope of the ordinate obtained in step 3 during planning, cost function adopts the gradient in band direction, column direction allows step to a certain degree;

Step 5. obtains up-and-down boundary and also just determines area-of-interest place, according to the candidate's bounding box obtained in this region screening step 3, extracts the HOG feature of residue candidate bounding box, and utilizes the SVM classifier of pre-training to determine whether pedestrian;

Step 6. utilizes the bounding box being judged to be pedestrian in yardstick and distance sorting procedure 5;

Final pedestrian's detection and tracking effect is obtained after step 7. is level and smooth in time by the result of step 6;

Step 8. does corresponding display and warning according to step 7 result.