SCENE OBSTRUCTION DETECTION USING HIGH PASS FILTERS

[0001] This application claims priority under 35 U.S.C 119(e)(1) to Provisional Application No. 62274525 filed Jan. 4, 2016.

[0002] The technical field of this invention is image processing, particularly to detect if the view of a fixed focus camera lens is obstructed by surface deposits (dust, road dirt, etc).

[0003] The fixed focus cameras used for Advanced Driver Assistance Systems (ADAS) are subject to many external conditions that may make the lens dirty from time to time. Car manufacturers are starting to design intelligent self-cleaning cameras that can detect dirt and automatically clean the lens using air or water.

[0004] One of the difficulties encountered in the prior art is the reliable detection of foreign objects such as dust, road dirt, snow, etc., obscuring the lens while ignoring large objects that are part of the scene being viewed by the cameras.

[0005] The solution shown applies to fixed focus cameras, widely used in automotive for ADAS applications. The problem solved by this invention is distinguishing a scene obscured by an obstruction, such as illustrated in FIG. 1, from a scene having large homogeneous areas, such as illustrated in FIG. 2. In accordance with this invention the distinction is made based upon the picture data produced by the camera. Obstructions created by deposits on a lens surface, as shown in FIG. 1, will appear blurred and will have predominantly low frequency content. A high pass filter may therefore be used to detect the obstructions.

[0006] A machine-learning algorithm is used to implement classification of the scene in this invention.

[0015] The steps required to implement the invention are shown in FIG. 3. The input image is first divided into a grid of N×M blocks in step 301. FIG. 4 illustrates the scene of FIG. 2 divided into a 3×3 set of blocks.

[0016] In step 302 the high frequency content of each block is computed by using horizontal and vertical high pass filters. This produces a total of 2×M×N values.

[0017] The reason for separately processing 3×3 (9) different regions of the image instead of the entire image is to calculate the standard deviation of the values across the image. The classified of this invention uses both mean and standard deviation values. Employing only the mean value could be sufficient to detect scenarios where the entire view is blocked but cannot prevent false-positive cases where one part of the image is obstructed and other parts are perfectly fine. The mean value cannot measure the high-frequency's contrast between different regions whereas the standard deviation can.

[0018] Step 303 then calculates the mean and the standard deviation for each high pass filter, across M×N values to form a 4 dimensional feature vector. Step 304 is an optional step that may augment the features vector an additional P component. This additional component may be meta information such as image brightness, temporal differences, etc.

[0019] Step 305 then classifies the scene as obscured or not obscured using a logistic regression algorithm having the feature vector as its input. This algorithm is well suited for binary classifications such as pass/fail, win/lose, or in this case blocked/not blocked.

[0020] This algorithm performs well where the two classes can be separated by a decision boundary in the form of a linear equation. Classification is shown in FIG. 5, where:

[0021] If θ₀+θ₁·x₁+θ₂·x_2≧0

• • [0022] then the (x₁,x₂) sample belongs to the X class 501 (image blocked) illustrated in FIG. 5,
    and

[0023] If θ₀+θ₁·x₁+θ₂·x₂<0

• • [0024] then the (x₁,x₂) sample belongs to the O class 502 (image clear) illustrated in FIG. 5.

[0025] In this invention the line is parametrized by θ=[θ₀, θ₁, θ₂] since the feature vector has two components x₁ and x₂. The task of the logistic regression is to find the optimal θ, which will minimize the classification error for the images used for training. In the case of scene obstruction detection, the feature vectors have 4 components [x₁, x₂, x₃, x₄] and thus the decision boundary is in form of an hyperplane with parameters [θ₀, θ₁, θ₂, θ₃, θ₄].

[0026] The training algorithm determines the parameter θ=[θ₀,θ₁,θ₂ . . . ] by performing the following tasks:

[0027] Gather all feature vectors into a matrix X and the corresponding classes into a vector Y.

$\begin{array}{cc}\begin{array}{c}X=\left\{\begin{array}{cccc}{X}_1^0& X_1^1& \dots & X_1^{M-1}\\ X_2^0& X_2^1& \dots & X_2^{M-1}\\ X_3^0& X_3^1& \dots & X_3^{M-1}\\ X_4^0& X_4^1& \dots & X_4^{M-1}\end{array}\right\}\\ =\left[X^0X^1\dots X^{M-1}\right]\end{array}& =\left[X^0X^1\dots X^{M-1}\right]\\ Y=\left[y^0y^1\dots y^{M-1}\right]\mathrm{where}y^k\mathrm{is}0\mathrm{or}1.& \end{array}$

[0028] Find θ=[θ₀, θ₁, θ₂, θ₃, θ₄] that minimizes the cost function:

$\begin{array}{c}J\left(\Theta \right)=\frac{1}{M}\sum _{k=0}^{M-1}\mathrm{Cost}\left(h_{\Theta }\left(X^k\right),y^k\right)\\ \mathrm{with}\text{:}\mathrm{Cost}\left(h_{\Theta }\left(X^k\right),y^k\right)=-y^k\log \left(h_{\Theta }\left(X^k\right)\right)-\left(1-y^k\right)\log \left(1-h_{\Theta }\left(X^k\right)\right)\mathrm{and}h_{\Theta }\left(X^k\right)=\frac{1}{1+e^{-{\Theta }^Tx^k}}\end{array}$

[0029] FIG. 6 shows the graphical representation of a sample cost function J(θ) for the case of a one dimensional feature vector.

[0030] Gradient descent is one of the techniques to find the optimum θmin which minimizes J(θ).

[0031] If for θmin we have Jθmin=0, this means the error rate for the classifier, when applied to the training data set, is 0%. However most of the time J(θmin)>0, which means there is some miss-classification error that can be quantified.

[0032] Next the algorithm's miss-classification error (also called accuracy) is calculated by applying the classifier rule to every feature vector of the dataset and comparing the results with the true result.

[0033] The final classification is done as follows:

[0034] If θ₀+θ₁·x₁+θ₂·x₂≧0

• • [0035] then the image is blocked,
    and

[0036] If θ₀+θ₁·x₁+θ₂·x₂<0

• • [0037] then the image is clear.

[0038] FIG. 7 illustrates an example system-on-chip (SOC) 700 suitable for this invention. SOC 700 includes general purpose central processing unit (CPU) 701, digital signal processor (DSP) 702, graphics processing unit (GPU) 703, video input ports 704, internal memory 705, display controller subsystem 706, peripherals 707 and external memory controller 708. In this example, all these parts are bidirectionally connected to a system bus 709. General purpose central processing unit 701 typically executes what is called control code. Control code is what gives SOC 700 its essential character generally in the way it interacts with the user. Thus CPU 701 controls how SOC 700 responds to user inputs (typically received via peripherals 707). DSP 702 typically operates to process images and real-time data. These processes are typically known as filtering. The processes FIG. 3 are performed by DSP 702. GPU 703 performs image synthesis and display oriented operations used for manipulation of the data to be displayed. Video input ports 704 receive the input images from possibly plural cameras. Video input ports 704 typically also includes suitable buffering of the image data prior to processing. Internal memory 705 stores data used by other units and may be used to pass data between units. The existence of memory 705 on SOC 700 does not preclude the possibility that CPU 701, DSP 702 and GPU 703 may include instruction and data cache. Display controller subsystem 706 generates the signals necessary to drive the external display used by the system. Peripherals 707 may include various parts such as a direct memory access controller, power control logic, programmable timers and external communication ports for exchange of data with external systems (as illustrated schematically in FIG. 7). External memory controller 708 controls data movement into and out of external memory 710.

[0039] A typical embodiment of this invention would include non-volatile memory as a part of external memory 710. The instructions to control SOC 700 to practice this invention are stored the non-volatile memory part of external memory 710. As an alternate, these instruction could be permanently stored in non-volatile memory part of external memory 710.