CN103942565B - Based on the quick object detecting method of iteration two points of cascade classifiers - Google Patents

Abstract

The present invention relates to a kind of quick object detecting method based on iteration two points of cascade classifiers, first, obtain initial strong classifier by AdaBoost algorithm; Secondly, with minimization calculation consumption for objective function, the strong classification that continuous iteration two points is initial, stop iteration when the absolute value of the calculating consumption difference of sorter in adjacent twice iteration two points of processes is less than given threshold value, the cascade classifier that now bipartite shape becomes is one and calculates the minimum global optimum's cascade classifier of consumption; Finally, this cascade classifier is utilized to carry out object detection in an image or a video.The object detecting method that the present invention proposes, can ensure, under the prerequisite that the detection perform of object detecting system is constant, by minimization calculation consumption, to realize quick object detection.

Description

Fast object detection method based on iterative binary cascade classifier

Technical field

The invention relates to a fast and efficient object detection method in the fields of human-computer interaction, computer vision, etc., in particular to a method for object detection using cascaded classifiers.

Background technique

Object detection is a very important research field in computer vision, including face detection, pedestrian detection and vehicle detection, etc. It can be widely used in human-computer interaction, video surveillance and image retrieval and other fields. The two main indicators to measure the quality of an object detection system are: detection rate and detection speed [1]. Generally speaking, a higher detection rate means a relatively slow detection speed, and a faster detection speed means a relatively low detection rate. Therefore, how to weigh the pros and cons of the two has always been an unavoidable problem in the field of object detection.

In recent years, with the development of smartphones and wearable devices, object detection based on mobile devices has gradually developed. However, due to limitations such as relatively weak computing power and relatively low battery power, mobile devices impose more stringent requirements on the real-time performance of object detection systems. Therefore, the present invention mainly studies how to improve the detection speed while keeping the detection rate unchanged.

Object detection mainly includes three aspects: feature extraction, window generation, and classifier judgment. Among them, the classifier decision occupies most of the object detection time. Researchers have done a lot of related work on the design of classifiers based on cascade structure, trying to reduce the time of classifier judgment by optimizing the structure of cascade classifiers, thereby speeding up the speed of object detection.

At present, most of the existing cascade classifier learning methods belong to the cascade classifier learning method based on detection rate and false detection rate, which is called DF-guided method for short. In 2004, Viola and Jones [2] found that the Boosted classifier with a relatively simple structure can reject most of the negative example windows on the premise of ensuring that all positive example windows pass. They use this feature to evenly distribute the total detection target to each classifier, that is, specify the detection rate and false detection rate of each classifier, and thus train a cascade classifier. This approach is known as the traditional cascade classifier learning approach. Since the first few stages of classifiers are only composed of a small number of weak classifiers and can reject most of the negative example windows in advance, the classifiers of this cascaded structure greatly speed up the speed of object detection. In 2008, Brubaker et al. [3] took advantage of the redundancy between weak classifiers at different levels, and proposed that the latter classifier can be continuously trained by using the scores of the previous classifier. This approach is known as the recycling cascade classifier approach. Compared with traditional cascaded classifiers, recycle cascaded classifiers reduces the number of weak classifiers in each classifier due to the reuse of information from previous classifiers, thus further speeding up the detection speed. In 2005, Bourdev and Brandt [4] proposed soft-Cascade. This method trains a strong classifier with a length of T, and sets a threshold for each weak classifier in the strong classifier, thus forming a cascaded classifier with a length of T. If a window passes the sum of the scores of the first t weak classifiers below the threshold of the tth weak classifier, it will be rejected immediately. This method reduces the total number of weak classifiers in cascaded classifiers, and by properly setting the threshold of each level of weak classifiers, the detection speed can be accelerated while the detection rate remains basically unchanged. The above several methods are proposed based on the idea of how to reduce the total number of weak classifications in each level and how to reject the negative window earlier. Although they have improved the detection speed to a certain extent, these methods have not fundamentally solved how to set the number of cascaded classifiers, how to allocate the detection rate and false detection rate of each classifier, and how to minimize the calculation consumption etc.

Compared with the DF-guided method, in recent years, researchers have begun to design cascaded classifiers from the perspective of minimizing the amount of calculation. In 2005, Chen and Yuille [5] selected weak classifications and generated cascade structures from the perspective of optimizing the total detection time. This method tentatively sets a larger total detection time and reduces the time from high to low until the time cannot be allocated to each level. At this time, the cascade classifier formed at this time is a fast, Efficient cascaded classifiers. This method speeds up the previous text detection algorithm [6] by 2.5 times. In 2010, Sabrian and Vasconcelos[7] started from the point of view that the design process of traditional cascaded classifiers did not consider the optimal speed and automatic design, and took the joint optimal classification error and calculation time as the objective function, and iteratively increased the maximum classifier during the training process. Weak classification capable of optimizing the objective function, a fast cascaded classifier (i.e. FCBoost) generation method is proposed. Compared with traditional cascaded classifiers, this method has a certain improvement in detection speed and detection performance. Similarly, in 2012, from the perspective of optimizing detection performance and computing speed, Chen[8] et al. continuously adjusted the sequence of weak classifiers, designed a Cronus cascade classifier and achieved good results. The above methods all design cascade classifiers from the perspective of lower computational complexity. Compared with the DF-guided method, they have achieved good results in detection speed and detection performance. However, most of the methods have problems such as training is too complex and local greedy.

references:

[1] G. Gualdi, A. Prati, amd R. Cucchiara. Multistage Particle Windows for Fast and Accurate Object Detection [J]. IEEE Transcationson Pattern Analysis and Machine Intelligence, 2012, 34(8): 1589-1604.

[2] P.Viola and M. Jones. Robust Real-Time Face Detection [J]. International Journal of Computer Vision, 2004, 57(2): 137-154.

[3] S. Brubaker, J. Wu, J. Sun, M. Mullin, and J. Regh. On the Design of Cascades of Boosted Ensembles for Face Detection [J]. International Journal of Computer Vision, 2008, 77 (1-3): 65-86.

[4] L. Bourdevand J. Brandt. Robust Object Detection via Soft Cascade [C]. In Proceedings of IEEE International conference on Computer Vision and Pattern Recognition, 2005.

[5] X. Chen and A. Yuille. A Time-Efficient Cascade for Real-Time Object Detection: With Applications for the Visually Impaired [C]. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition, 2005.

[6] X. Chen and A. L. Yuille. Detecting and Reading Text in Natural Scenes [C]. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2004.

[7] M. Saberian and N. Vasconcelos. Boosting Classifier Cascades [C]. In Proceedings of Advances in Neural Information Processing Systems, 2010.

[8] M. Chen, Z. Xu, K. Weinberger, O. Chapelle, and D. Kedem. Classifier Cascade for Minimizing Feature Evaluation Cost Minmin [C]. In Proceedings of International Conference on Artificial Intelligence and Statistics, 2012.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies such as complex training and local greed in the design of cascaded classifiers in the existing object detection process, and propose a fast object detection method. The object detection method proposed in the present invention can ensure the accuracy of the object detection system. Under the premise of the same detection performance, fast object detection is realized by minimizing the calculation consumption. Technical scheme of the present invention is as follows:

A fast object detection method based on an iterative binary cascade classifier. First, the initial strong classifier is obtained through the AdaBoost algorithm; secondly, with the objective function of minimizing the calculation consumption, the initial strong classification of the two points is continuously iterated. When adjacent Stop the iteration when the absolute value of the difference in the calculation consumption of the classifier during the two iterations of the binary division is less than a given threshold, and the cascade classifier formed by the bisection at this time is a globally optimal cascade classifier with the smallest calculation consumption; Finally, use this cascaded classifier for object detection in images or videos.

Specifically include the following steps:

Step 1: Collect a large number of positive samples and negative samples about the detected object, and set the performance that needs to be achieved in the training process: detection rate and false detection rate.

Step 2: Using the above positive samples, negative samples, detection rate and false detection rate, use the AdaBoost algorithm to train an initial strong classifier composed of T weak classifiers and its classification threshold t, where x represents positive and negative samples, h _i (x) represents the i-th weak classifier, and α _i represents the weight of the i-th weak classifier;

Step 3: According to the above positive samples and classification threshold t, sequentially calculate the maximum value M(r) of the sum of the response values of the last Tr weak classifiers in the strong classifier, r=1,...,T-1 ; Then calculate the difference between the classification threshold classification threshold t of the strong classifier H(x) and the corresponding maximum value M(r) respectively, and obtain the classification threshold t _r when the weak classifier r dichotomizes the strong classifier H(x), That is, t _r =tM(r), r=1,...,T-1;

Step 4: Use the classification threshold _{t r} obtained above to find the optimal r ₁ , so that the calculation consumption f ₁ minimum;

Step 5: Fix r ₁ , seek an optimal r ₂ after r ₁ , and minimize the calculation consumption f ₂ of the three-level cascaded classifier formed by dichotomizing H(x) at the r ₂ weak classifier ;

Step 6: a) fix r ₂ , update r ₁ within the range of 1 to r ₂ , so as to minimize the calculation consumption f ₂ of the three-level cascaded classifier; b) fix r ₁ , and update r 1 within the range of r ₁ to T r ₂ , so that the calculation consumption f ₂ of the three-level cascade classifier is the smallest; c) Repeat the process a, b, iteratively update r ₁ , r ₂ , and stop the iterative update when the two no longer change;

Step 7: Fix r ₁ and r ₂ , and find an optimal weak classifier r ₃ after r ₂ , so as to minimize the calculation amount f ₃ of the four-level cascade classifier formed by dichotomizing at the r _3rd weak classifier; According to the idea of step 6, iteratively update r ₁ , r ₂ and r ₃ continuously, and stop the iterative update when the three no longer change;

Step 8: Set the variable i to 3, a) fix the previously obtained r ₁ ,…,r _i , seek an optimal r _i+1 , and make the cascade classifier formed by dichotomy at r _i+1 calculate the consumption Minimum, b) Continuously iteratively update r ₁ ,...,ri ,r _{i +1} according to step 6. When two adjacent iterations are updated, r ₁ ,...,ri ,r _{i +1} will no longer occur Stop the iterative update when it changes; c) If the absolute value Δf of the difference in calculation consumption before and after increasing r _i+1 is less than the given threshold Δ, stop dividing H(x), otherwise, add 1 to i and continue the process of step 8 a;

Step 9: The cascade classifier of level i+1 obtained by dichotomizing r ₁ ,...,r _i is a globally optimal cascade classifier with the least computational consumption;

Step 10: Use the cascade classifier obtained in step 9 to perform object detection in images or videos.

Among them, in step 4, the method of seeking the optimal r ₁ to minimize the calculation consumption f ₁ of the second-level cascade classifier formed by the dichotomous strong classifier H(x) at the r _1th weak classifier is: Among them, p represents the rejection rate of negative samples, that is, the percentage of the number of negative samples rejected by the first r weak classifiers to the total number of negative samples.

Using the method of the present invention, the cascade classifier obtained by continuously iterating the binary original strong classifier H(x) is a globally optimal cascade classifier based on the minimization of calculation consumption. Compared with the object detection method based on the traditional cascade classifier, the object detection method based on the iterative binary cascade classifier effectively reduces the average number of features used in each window, thereby reducing the computational consumption of the classifier, Speed up object detection. At the same time, the method is simple, unlike the traditional cascade classifier design method, which needs to set the number of cascade classifiers and assign the detection rate and false detection rate of each level.

Description of drawings

Fig. 1 is a block diagram of the proposed method of the present invention.

detailed description

Below in conjunction with accompanying drawing and the present invention is described:

The present invention assumes that all weak classifications have the same amount of computation, and all of them are 1. Let H(x) be a strong classifier trained by AdaBoost algorithm, and t represents the classification threshold of the strong classifier, then it can be expressed as

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Among them, x represents the detection window, h _i (x) represents the i-th weak classifier, α _i represents the weight of the i-th weak classifier, and T represents the total number of weak classifiers in the strong classifier. Then, when the response value H(x) of the detection window x is greater than the threshold t of the given classifier, the window is a positive example window; otherwise, the window is a negative example window.

Assuming that H(x) is divided into two parts at the rth weak classifier, H _L (x) and _HR (x), it is expressed as

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; R</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; r</span>}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Assuming we have a large number of detection windows x, then for each weak classifier r, the maximum value maxH _R (x, r) of its HR ₍ x, r) can be calculated, denoted as M(r).

If a detection window x satisfies inequality (3)

H _L (x, r) + M (r) ≤ t, (3)

Then the detection window x can be directly judged as a negative example window without calculating the response values of the remaining Tr weak classifications. At this time, _HL (x, r) can be regarded as a first-level classifier, and its classification threshold is _tM ( _r ), and HL (x, r) and HR (x, r) can be regarded as a second-level classifier together , and its classification threshold is t. If the response value of the detection window passed by the first r weak classifiers is less than tM(r), it is directly judged as a negative example window; if the response value of the first r weak classifiers is greater than tM(r), the detection window needs further Calculate the response values of the remaining Tr weak classifiers. If the sum of the response values of these T weak classifiers is greater than the classification threshold t, it is judged as a positive window, otherwise it is a negative window. It can be seen that, because the first r weak classifiers reject some negative example windows, compared with the first-level strong classifier, the second-level cascaded classifier reduces the calculation consumption and speeds up the detection speed.

The present invention starts from the idea of whether the detection window can be rejected earlier with fewer weak classifiers, and proposes a fast object detection method based on iterative binary cascading classifiers. In this method, the objective function is to minimize the calculation consumption, and iteratively iterates the binary original one-level strong classifier H(x) until the calculation consumption f converges. Referring to Figure 1, the specific steps are as follows:

Step 1: Collect a large number of positive samples and negative samples about the detected object, and set the performance that needs to be achieved in the training process: detection rate and false detection rate;

Step 2: Using the above positive and negative samples, detection rate and false detection rate, use the AdaBoost algorithm to train a strong classifier composed of T weak classifiers and its classification threshold t, where x represents positive and negative samples, h _i (x) represents the i-th weak classifier, and α _i represents the weight of the i-th weak classifier.

Step 3: According to the above positive sample set, sequentially calculate the maximum value M(r) of the sum of the response values of the last Tr weak classifiers in the strong classifier, r=1,...,T-1; and then respectively Calculate the difference between the threshold t of the strong classifier H(x) and these maximum values M(r), and obtain the classification threshold t _r when the weak classifier r dichotomizes the strong classifier H(x), that is, t _r =tM(r), r=1,...,T-1.

Step 4: Use the classification threshold _{t r} obtained above to find the optimal r ₁ , so that the calculation consumption f ₁ minimum, i.e.

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\underset{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\underset{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; pr</span>}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

Among them, p represents the rejection rate of negative samples, that is, the percentage of the number of negative samples rejected by the first r weak classifiers to the total number of negative samples.

Step 5: Fix r ₁ , seek an optimal r ₂ after r ₁ , and minimize the calculation consumption f ₂ of the three-level cascaded classifier formed by dichotomizing H(x) at the r ₂ weak classifier .

Step 6: a) fix r ₂ , update r ₁ within the range of 1 to r ₂ , so as to minimize the calculation consumption f ₂ of the three-level cascaded classifier; b) fix r ₁ , and update r 1 within the range of r ₁ to T r ₂ , to minimize the calculation consumption f ₂ of the three-level cascaded classifier; c) Repeat the process a and b, iteratively update r ₁ and r ₂ , and stop the iterative update when the two no longer change.

Step 7: Fix r ₁ and r ₂ , and find an optimal weak classifier r ₃ after r ₂ , so as to minimize the calculation amount f ₃ of the four-level cascade classifier formed by dichotomizing at the r _3rd weak classifier; According to the idea of step 6, iteratively update r ₁ , r ₂ and r ₃ continuously, and stop the iterative update when the three no longer change;

Step 8: Set the variable i to 3, a) fix the previously obtained r ₁ ,…,r _i , seek an optimal r _i+1 , and make the cascade classifier formed by dichotomy at r _i+1 calculate the consumption Minimum, b) Continuously iteratively update r ₁ ,...,ri ,r _{i +1} according to step 6. When two adjacent iterations are updated, r ₁ ,...,ri ,r _{i +1} will no longer occur Stop the iterative update when it changes; c) If the absolute value Δf of the difference in calculation consumption before and after increasing r _i+1 is less than the given threshold Δ, stop dividing H(x), otherwise, add 1 to i and continue the process of step 8 a.

Step 9: The i+1 level cascade classifier obtained by dichotomizing r ₁ ,...,r _i is a globally optimal cascade classifier with the least computational consumption.

Step 10: Use the cascade classifier obtained in step 9 to perform object detection in images or videos.

Claims (2)

1. based on a quick object detecting method for iteration two points of cascade classifiers, first, obtain initial strong classifier by AdaBoost algorithm; Secondly, with minimization calculation consumption for objective function, the strong classification that continuous iteration two points is initial, stop iteration when the absolute value of the calculating consumption difference of sorter in adjacent twice iteration two points of processes is less than given threshold value, the cascade classifier that now bipartite shape becomes is one and calculates the minimum global optimum's cascade classifier of consumption; Finally, utilize this cascade classifier to carry out object detection in an image or a video, specifically comprise the following steps:

Step 1: collect the positive example sample of a large amount of related detection object and negative routine sample, and the performance setting that training process needs to reach: verification and measurement ratio and empty inspection rate;

Step 2: utilize above-mentioned positive example sample, negative routine sample and verification and measurement ratio and empty inspection rate, use AdaBoost Algorithm for Training to obtain an initial strong classifier be made up of T Weak Classifier and classification thresholds t, wherein, x represents positive and negative routine sample, h _ix () represents i-th Weak Classifier, α _irepresent the weight of i-th Weak Classifier;

Step 3: according to above-mentioned positive example sample and classification thresholds t, calculates the maximal value M (r) of the response sum of T-r Weak Classifier after in strong classifier, r=1 successively ..., T-1; Then calculate classification thresholds classification thresholds t and the difference of maximal value M (r) corresponding separately of strong classifier H (x) respectively, obtain classification thresholds t during Weak Classifier r bis-points of strong classifiers H (x) _r, i.e. t _r=t-M (r), r=1 ..., T-1;

Step 4: utilize classification thresholds t obtained above _r, seek optimum r ₁, make at r ₁the calculating consumption f of the two-level concatenation sorter that two points, individual Weak Classifier place strong classifier H (x) is formed ₁minimum;

Step 5: fixing r ₁, at r ₁seek an optimum r afterwards ₂, make at r ₂individual Weak Classifier place continues the calculating consumption f of the three-stage cascade sorter that two points of H (x) are formed ₂minimum;

Step 6:a) fixing r ₂, 1 to r ₂r is upgraded in scope ₁, make the calculating consumption f of three-stage cascade sorter ₂minimum; B) fixing r ₁, at r ₁r is upgraded in the scope of T ₂, make the calculating consumption f of three-stage cascade sorter ₂minimum; C) repetitive process a, b, continuous iteration upgrades r ₁, r ₂, stop iteration upgrading when the two no longer changes;

Step 7: fixing r ₁, r ₂, at r ₂find an optimum Weak Classifier r afterwards ₃, make at r ₃the calculated amount f of the level Four cascade classifier that individual Weak Classifier place bipartite shape becomes ₃minimum; According to the thought of step 6, continuous iteration upgrades r ₁, r ₂and r ₃, stop iteration upgrading when three no longer changes;

Step 8: set variable i as 3, the r obtained before a) fixing ₁..., r _i, seek an optimum r _i+1, make at r _i+1it is minimum that the cascade classifier that place's bipartite shape becomes calculates consumption, b) upgrades r according to the continuous iteration of the mode of step 6 ₁..., r _i, r _i+1, in adjacent twice iteration upgrades, r ₁..., r _i, r _i+1iteration is stopped to upgrade when all no longer changing; If c) increase r _i+1when the absolute value delta f of the difference of the calculating consumption of front and back is less than given threshold value Δ, stops two points of H (x), otherwise i is added 1, continue the process of step 8 a);

Step 9: by r ₁..., r _itwo points obtain i+1 level cascade classifier and are one and calculate the minimum global optimum's cascade classifier of consumption;

Step 10: the cascade classifier utilizing step 9 to obtain carries out object detection in an image or a video.

2. quick object detecting method according to claim 1, is characterized in that, in step 4, seeks optimum r ₁, make at r ₁the calculating consumption f of the two-level concatenation sorter that two points, individual Weak Classifier place strong classifier H (x) is formed ₁minimum method is: wherein, p represents the reject rate of negative routine sample, is namely accounted for the number percent of negative routine total sample number by the negative routine number of samples of front r Weak Classifier refusal.