US20150324659A1

US20150324659A1 - Method for detecting objects in stereo images

Info

Publication number: US20150324659A1
Application number: US14/272,570
Authority: US
Inventors: Ming-Yu Liu; Oncel Tuzel
Original assignee: Mitsubishi Electric Research Laboratories Inc
Current assignee: Mitsubishi Electric Research Laboratories Inc
Priority date: 2014-05-08
Filing date: 2014-05-08
Publication date: 2015-11-12
Anticipated expiration: 2034-05-08
Also published as: JP2015215877A; JP6345147B2; US9195904B1; CN105096307B; CN105096307A

Abstract

A method detects an object in a pair of stereo images acquired of a scene, by first generating a cost volume from the pair of stereo images, wherein the cost volume includes matching costs for a range of disparity values, for each pixel the sterao images, between the stereo images in the pair. Feature vectors are determined from sub-images in the cost volume using a feature function of the disparity values with a minimal accumulated cost within regions inside the sub-images. Then, a classifier is appled to the feature vectors to detect whether the sub-image includes the object.

Description

FIELD OF THE INVENTION

This invention relates to computer vision, and more particularly to detecting objects in stereo images.

BACKGROUND OF THE INVENTION

Many computer vision applications use stereo images acquired by a stereo camera to detect objects. A stereo camera typically has multiple lenses and sensors. Usually, the intra-axial distance between the lenses is about the same distance as between the eyes to provide overlapping views.
FIG. 1 shows a conventional system for stereo-based object detection. A stereo camera 101 acquires stereo images 102. The dection method can include the following steps: stereo imaging 100, cost volume determination 110, depth/disparity map estimation 120, and object detection 130.
Most of the conventional methods for stereo-based object detection rely on per pixel depth information in the overlapping area 120. This step is generally referred as depth/range map estimation. This step can be achieved by determining disparity values, i.e., translation of corresponding pixels in the two images, determining the depth map. The depth map can then be used lot object detection 130, e.g., a histogram of oriented gradients (HoG) of the depth map is used for object description. One method estimates the dominate disparity in a sub-image region, and use a co-occurrence histogram of the relative disparity values for object detection.
Depth/range/disparity map estimation is a challenging problem. Local methods suffers from inaccurate depth determination, while global methods require significant computational resources, and are unsuited for real-time applications.
Several methods avoid the depth map determination step by using stereo cues for region of interest generation. For example, one method determines a stixel map which marks the potential object locations. Each stixel is defined by a 3D position relative to the camera and stands vertically on a ground plane. A detector based on the color image content is then applied to the locations to detect objects.
U.S. Publication 20130177237 uses a range map to determine an area of interest, and uses a classifier based on an intensity histogram to detect objects.
Region of interest methods cannot be directly applied to object detection. They have to be applied in conjunction with other object detectors. In addition, miss detection is certain when the area of interest does not cover the object.

SUMMARY OF THE INVENTION

The embodiments of the invention provide a method for detecting objects in stereo images. A cost volume is computed from the images. Then, object detection is directly applied features obtained from the cost volume. The detection uses T decision tree classifiers (Adaboost) that are learned from training features.
The invention avoids the error-prone and computationally-complex depth map estimation step of the prior art, and leads to an accurate and efficient object detector. The method is better-suited for embedded systems because it does not require complex optimization modules necessary to obtain good depth map. In addition, the method searches all sub-images in the input images to detect the object. This avoids the miss detection problem that exists in the region of interest generation techniques.
The detection is accurate because the method can leverage a large amount of training data and make use of machine learning procedures. It outperforms region of interest generation techniques in detection accuracy.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block diagram of a conventional stereo-based object detection system:

FIG. 2 is a block diagram of a stereo-based object detection system according to embodiments of the invention;

FIG. 3 is a block, diagram of an object detection module for the stereo-based object detection system of FIG. 2;

FIG. 4 is a block diagram of a method for learning the stereo-based object detector according to embodiments of the invention.

FIG. 5 is a schematic of cost volume determination according to embodiments of the invention;

FIG. 6 is a schematic of a learned feature according to embodiments of the invention; and

FIG. 7 is a schematic of objects occupying large and small portions of sub-images.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

FIG. 2 shows a method and system for detecting an object 201 in a pair al stereo images 200 according, to embodiments of our invention. A cost volume 211 is generated 210 from the pair of stereo images. This is followed by selecting and extracting 215 feature vectors 216. Then, an object detector 220 is applied to the features vectors volume to detect the object. The object detector uses classifiers 230 leaned form training images features 231. After the object is dected, it can be localized, that is the loation of the object in the image can be determined. The method can be performed in a processor 250 connected to memory and input/output interfaces by buses as known in the art.
Our invention is based on the realization that depth information available in a depth map is also available in the cost volume, because the depth map is derived from the cost volume.
Our detector 220 that uses the cost volume directly is theoretically capable of matching the performance of any detector based on the depth map. Moreover, the cost volume is a richer representation than the conventional depth map. The depth map only provides a depth for each pixel, while the cost volume provides matching costs for a range of potential depths that each pixel in the stereo images can have, including the true depth. Hence, the detector uses features directly obtained from the cost volume can access more depth information, and achieve better performance.
As shown in FIG. 3, one embodiment of our invention includes cost volume generation 210, feature extraction 310, object detection and localization 320, learned discriminative features 330, and a learned object classification model 340. The localization determines where the object is detected.
FIG. 4 shows a machine learning procedure for learning the discriminative features and the learned object classification model. Features are selected and learned 410 from training data 400 comprising pairs of training stereo images.
Cost Volume Generation
FIG. 5 shows the generation of the cost volume C 211. The cost volume C: X×Y×D is a three-dimension data structure stored in the memory, where X and Y denote the image x and y axes and D denotes a set of disparity values, which are translations between corresponding pixels in the two stereo images I_L 501 and I_R 502. We assume that L_Land I_Rare rectified, which means that the images have been transformed such that the lens distortion effects are compensated, and a pixel in a row of one image is mapped to a pixel in the same row of the other image. The cost volume can then be determined by matching pixel appearance in the pair of stereo images I_Land I_R.
One way to determine the cost volume apply the mapping given by
C(x, y, d) =∥I _L(x, y)−I_R(x−d,y)∥₂+λ∥
grad(I_L(x,y))−grad(I_R(x−d,y))∥₂
for any (x,y,d)∈X×Y×D.
where ∥∥₂denotes a Euclidean norm and I_L(x, y) refers the pixel color values in the (x,y) location of the I_Limage, I_R(x,y) refers the pixel color values in the (x,y) location of the IR image, grad(I_L(x,y)) refers to the gradient in the (x,y) location of the IL image, grad(I_R(x−d,y)) refers to the gradient in the (x,y) location of the IR image, and λ is the weight controlling the importance of the gradient information. Note that image smoothing technique can be applied, such as bilateral filtering or guided filtering to enhance the cost volume.
Feature Extraction
FIG. 6 shows feature selection and extraction 215 of FIG. 2. We extract a K-dimensional feature vector from each sub-image 600 for determining whether or not the object is present in the sub-image. The sub-images can be considered a moving window passed over the image in, e.g., in a raster scan order for eachy pixel at multiple scales.
Note, the embodiments only directly uses the cost volume to determine the features. Depth map estimation as in the prior art is not performed.
Each dimension of the feature vector corresponds to a numerical comparison result between a Fig cost disparity values of two e.g., rectangle, regions R_k ¹ 601 and R _k ² 602 in the sub-image 600. Let the sub-image be denoted as J and the k^thdimension of feature vector be represented as f_k(J). The value of f_k(J) is
$\begin{matrix} f_{k} (J) = {\begin{matrix} 1 & if d_{\min} (R_{k}^{1}) > d_{\min} (R_{k}^{2}) \\ 0 & if d_{\min} (R_{k}^{1}) = d_{\min} (R_{k}^{2}) \\ - 1 & otherwise \end{matrix}, & (1) \end{matrix}$
where d_min(R_k ⁱ) represents to the disparity value that has a minimal (min) accumulated cost in the region of R_k ⁱof the sub-image. That is
$\begin{matrix} d_{\min} (R_{k}^{i}) = \arg \min_{d} \sum_{(x, y) \in R_{k}^{i}}^{} C (x, y, d) . & (2) \end{matrix}$
Note that determining the minimal cost disparity value in the region is relatively simple because the accumulated cost can be obtained efficiently using an integral image technique as known in the art. The locations and size of the regions are learned using a machine learning procedure, which is described below.
Object Detection and Localization
The K-dimensional feature vector associated with the sub-image is passed to an ensemble classifier for determining a detection score. The ensemble classifier includes T decision tree classifiers. Each decision tree classifier takes a small number of dimensions of the K-dimensional feature as input, and classifies the sub-image as positive (containing an object) or negative (not containing an object). A detection score s obtained from the classifier for the sub-image J is given by
$\begin{matrix} s (J) = \sum_{t = 1}^{T} θ_{t} δ_{t} (J), & (3) \end{matrix}$
where δ_t's are the decision tree classifiers and θ_t's are the corresponding weights. If the score is greater than a preset threshold, then the system declares a detection in the sub-image.
As shown in FIG. 7, the classifier can be trained to give a higher score when the object occupies a larger portion, of the sub-image 701 and a lower score when the object only occupies a small portion of the subimage 702, because the larger object provides a better estimate of where the object is located within the image than the smaller object.
Feature Selection and Classifier Learning Procedure
We use a discrete AdaBoost procedure for selecting the region
{(R _k ¹ ,R _k ²)|∀k=12, . . . , K}, (4)
and for learning the decision tree classifier weights
{θ_t|∀t=1, 2, . . . , T}. (5).
We collect a set of data fir a learning task, which includes a set of stereo training images. The sub-images that contain an object is labeled as positive instances, while others are labeled as negative instances. We align the positive and negative sub-images so that their centers coincide. The sub-images are also scaled to have the same height. The aligned and scaled sub-images are denoted as
D={(j _i ,l _i), i=1, 2, . . . , V}, (6)
where J_idenotes the i^thsub-image, l_iis the label, and V is the total number of sub-images.
We sample a set of N regions as the feature pools {R_i, i=1, 2, . . . , N}, which have different locations and sizes and are covered by the aligned sub-images. We randomly pair two regions and compare their disparity values of the minimal cost. This is performed K times to construct a K-dimensional feature vector.
We use the discrete AdaBoost procedure to learn T decision tree classifiers and their weights. The procedure starts with assigning uniform weights to the training samples. A decision tree is then learned based on the current training sample weights. The weights of incorrectly classified samples are increased so that the weights have more impact during the next round of decision tree classifier learning. We assign the weight to the decision tree classifier based on the weighted error rate. This process is repeated I times to construct an ensemble classifier of T decision tree classifiers. A pseudo code of the procedure is described below.

Input: Feature vectors and class labels D={(f(J_i),l_i), i=1, 2, . . . , V}
Output: Ensemble classifiers Σ_t=1 ^Tθ_tδ_t(J)
Start with uniform weights

$w_{i} = \frac{1}{V}, i = 1, 2, \dots, V$
For t=1, 2, . . . , T

1. Learn a decision tree classifier δ_t(J)∈{−1,1} using weights w_i's;
2. Determine error rate ε=Σ_iw_i|(δ_t(J_i≠l_i);
3. Determine decision tree classifier weight

$θ_{t} = \log \frac{1 - ɛ}{ɛ};$

4. Set w_i←w_iexp(θ_t|(δ_t(J_i)≠l_i) for i=1, 2, . . . , V; and
5. Normalize the sample weights

$w_{i} \leftarrow \frac{w_{i}}{Σ_{i} w_{i} .}$
Function δt, which is used in steps 2 and 4, represents the indicator function, which returns one if the statement in the parenthesis is true and zero otherwise.
Although the invention has been described by way of examples of preferred embodiments, it is to be understood that various other adaptations and modifications can be made within the spirit and scope of the invention. Therefore, it is the object of the appended claims to cover all such variations and modifications as come within the true spirit and scope of the invention.

Claims

1. A method for detecting an object in a pair of stereo images acquired of a scene, where each stereo image include pixels, comprising the steps of:

generating a cost volume from the pair of stereo images, wherein the cost volume includes matching costs for a range of disparity values, for each pixel, between the stereo images in the pair;

determining feature vectors from sub-images in the cost volume using a feature function of the disparity values with a minimal accumulated cost within regions inside the sub-images; and

applying a classifier to the feature vectors to detect whether the sub-images includes the object, wherein the steps are performed in a processor.

2. The method of claim 1, further comprising:

localizing the object within the stereo images.

3. The method of claim 1, wherin the classifier is learned from pairs of training stereo images.

4. The method of claim 1, further comprising:

rectifying the pair of stereo images.

5. The method of claim 1, further comprising:

smoothing the pair of stereo images.

6. The method of claim 1, wherein the generating further comprises:

matching colors and gradients of the pixels in the pair of stereo images using a Euclidaian norm.

7. The method of claim 1, wherein the feature function is

\begin{matrix} f_{k} (J) = {\begin{matrix} 1 & if d_{\min} (R_{k}^{1}) > d_{\min} (R_{k}^{2}) \\ 0 & if d_{\min} (R_{k}^{1}) = d_{\min} (R_{k}^{2}) \\ - 1 & otherwise \end{matrix}, & (1) \end{matrix}

where J represents the sub-image, k represents a dimension of the feature vectors, min represents a function that returns a minimum, and d_min(R_k ⁱ) represents the disparity value that has a minimal accumulated cost in rectangle area of R_k ⁱin the sub-image, wherein i indexed the rectangular regions.

8. The method of claim 7, wherein

\begin{matrix} d_{\min} (R_{k}^{i}) = \arg \min_{d} \sum_{(x, y) \in R_{k}^{i}}^{} C (x, y, d) . \end{matrix}

where C(x, y, d) represent the cost volume.

9. The method of claim 1, wherein the classifier is an ensemble classifier including T decision tree classifiers.

10. 10. The method of claim 9, wherein the classifier provides a detection score s for sub-image J is

s (J) = \sum_{t = 1}^{T} θ_{t} δ_{t} (J),

where δ_t's are the decision tree classifiers and θ_t's are corresponding weights.