A method and apparatus for detecting objects of interest in images

The invention relates to a method and apparatus for detecting objects of interest in images and in particular to object de ^¬ tection using deep neural networks. The detection of objects is a common concern in various use cases. In use cases like surveillance or driver assistance systems or autonomous driving or any other use cases where information about the environment is required, the detection of objects is an important task.

Accordingly, it is an object of the present invention to pro ^¬ vide a method and apparatus for detecting objects of interest in images efficiently. This object is achieved according to a first aspect of the present invention by a method for detecting objects of interest in images comprising the features of claim 1.

The invention provides according to the first aspect a method for detecting objects of interest in images,

wherein the method comprises the steps of:

supplying at least one input image to a trained deep neural network DNN which comprises a stack of layers, and using a deconvolved output of a learned filter or combining

deconvolved outputs of learned filters of at least one layer of the trained deep neural network DNN to detect the objects of interest in the supplied images.

In a possible embodiment of the method according to the first aspect of the present invention, the deep neural network DNN is trained on annotated training images comprising annotation labels . In a further possible embodiment of the method according to the first aspect of the present invention, the annotation la ^¬ bels of the annotated training images comprise count values indicating a number of objects of interest within the respec- tive training image.

In a still further possible embodiment of the method accord ^¬ ing to the first aspect of the present invention, the deep neural network DNN is trained on the basis of annotated syn- thetic training images and/or annotated natural training im ^¬ ages .

In a still further possible embodiment of the method accord ^¬ ing to the first aspect of the present invention, deconvolved outputs of learned filters of mid-level and high-level layers of the trained deep neural network DNN are combined to detect the objects of interest in the supplied images.

In a further possible embodiment of the method according to the first aspect of the present invention, the input images comprise pixels with pixel values for different channels.

In a further possible embodiment of the method according to the first aspect of the present invention, the channels com- prise color channels including pixel values for different colors including a red color, a blue color and a green color.

In a further possible embodiment of the method according to the first aspect of the present invention, a noise within the at least one deconvolved output of at least one learned fil ^¬ ter is reduced by processing the different channels separate- ly.

In a further possible embodiment of the method according to the first aspect of the present invention, at least one chan ^¬ nel is preserved while at least another channel is suppressed to reduce a noise within at least one deconvolved output of at least one learned filter. In a still further possible embodiment of the method accord ^¬ ing to the first aspect of the present invention, selected channels are processed to increase a mean value of the re- spective channels.

In a further possible embodiment of the method according to the first aspect of the present invention, a background with ^¬ in the input image is eliminated at least partially by com- paring the output of a learned filter of a layer with the deconvolved outputs of learned filters.

In a further possible embodiment of the method according to the first aspect of the present invention, a thresholding is performed for detecting contours of objects of interest with ^¬ in the supplied images.

In a further possible embodiment of the method according to the first aspect of the present invention, bounding boxes are superimposed on the input images around the detected contours to mark objects of interest within the input images.

In a further possible embodiment of the method according to the first aspect of the present invention, the input images with the superimposed bounding boxes are memorized, transmit ^¬ ted and/or displayed and/or wherein the bounding boxes are stored separately as meta data.

In a still further possible embodiment of the method accord- ing to the first aspect of the present invention, the objects of interest comprise mobile objects including pedestrians and/or vehicles and/or static objects.

In a still further possible embodiment of the method accord- ing to the first aspect of the present invention, the deep neural network DNN comprises a convolutional neural network CNN and/or a recurrent neural network RNN and/or a deep feed forward network. The invention further provides according to a further aspect an apparatus for detection of objects of interest in images comprising the features of claim 15.

The invention provides according to the second aspect an ap ^¬ paratus for detection of objects of interest in images where ^¬ in the apparatus comprises:

a supply unit adapted to supply at least one input image to a trained deep neural network DNN which comprises a stack of layers; and

a processing unit adapted to combine deconvolved outputs of learned filters of at least one layer of the trained deep neural network DNN to detect objects of interest in the sup ^¬ plied input images.

In the following, possible embodiments of the different as ^¬ pects of the present invention are described in more detail with reference to the enclosed figures.

Fig. 1 shows a block diagram of a possible exemplary embodiment of an apparatus for detection of objects of interest in images according to an aspect of the present invention;

Fig. 2 shows a flowchart of a possible exemplary embodi ^¬ ment of a method for detection of objects of inter ^¬ est in images according to a further aspect of the present invention;

Fig. 3 shows a block diagram of a possible further exemplary embodiment of a detection apparatus according to the present invention;

Fig. 4 shows a schematic diagram for illustrating the operation of a method and apparatus for detecting ob ^¬ jects of interest in images according to a possible embodiment . As can be seen in Fig. 1, the detection apparatus 1 according to the present invention comprises in the illustrated embodi ^¬ ment a supply unit 2 and a processing unit 3. The supply unit 2 is adapted to supply at least one input image to a trained deep neural network DNN which comprises a stack of layers. The output of the supply unit 2 is connected to a processing unit 3 of the detection apparatus 1. In a possible embodiment the processing unit 3 is adapted to combine deconvolved out- puts of learned filters of at least one layer of the trained deep neural network DNN to detect objects of interest in the input images supplied to the trained deep neural network by the supply unit 2. The processing unit 3 can also be adapted to use a deconvolved output of a learned filter of a layer of the DNN.

The images supplied to the trained deep neural network DNN can be captured in a possible embodiment by cameras. In an alternative embodiment, the input images supplied to the trained deep neural network DNN can be read from a memory or can be received from other devices of a system.

The detection apparatus 1 is provided for detecting objects of interest in received digital images. The objects of inter- est can comprise either mobile objects or static objects. The mobile objects can be for instance pedestrians and/or vehi ^¬ cles in the supplied digital images. Further, the objects can also comprise static objects within the supplied digital im ^¬ ages such as trees or buildings.

The deep neural network DNN can be trained on annotated training images comprising annotation labels. In a possible embodiment, the annotation labels of the annotated training images can comprise count values indicating a number of ob- jects of interest with the respective training image supplied to the deep neural network DNN to be trained. The deep neural network DNN can be trained on the basis of annotated synthet- ic training images and also on annotated natural training im ^¬ ages .

The deep neural network DNN comprises a stack of layers.

Deconvolved outputs of learned filters of mid-level and high- level layers of the trained deep neural network DNN are com ^¬ bined in a possible embodiment to detect the objects of in ^¬ terest in the supplied digital images. In a possible embodiment, the input images supplied to the trained deep neural network DNN can comprise pixels wherein each pixel comprises pixel values. In a possible implementa ^¬ tion, a pixel can comprise pixel values for different chan ^¬ nels. In a possible embodiment, these channels comprise color channels including pixel values for different colors. In a possible implementation, the color channels can comprise pix ^¬ el values for a red color channel, a blue color channel and a green color channel. In a possible embodiment, a noise within the combined

deconvolved outputs of the learned filters can be reduced by processing the different channels separately. In a possible embodiment, at least one channel such as a color channel is preserved while at least another channel such as another col- or channel can be suppressed to reduce a noise within the combined deconvolved outputs of the learned filters. The se ^¬ lected channels such as color channels can also be processed to increase a mean value of a respective channel. In a possible embodiment of the detection apparatus 1, a background within the input image can be eliminated at least partially by comparing the output of a learned filter of a layer with the deconvolved output of a learned filter or a combination of the deconvolved outputs of learned filters of the deep neural network DNN. Then, a thresholding can be performed for detecting contours of objects of interest within the supplied digital images. Further, the detection apparatus 1 can be adapted to superimpose bounding boxes on the input images around the detected contours to mark objects of interest within the input images. In a possible embodiment, the input images with the superim- posed bounding boxes can be memorized as meta data in a data memory. Further, thebounding boxes corresponding to the input images can also be memorized, transmitted to other devices of the system and/or stored separately as meta data. Moreover, the input images with the superimposed bounding boxes can al ^¬ so be displayed on a screen to a user.

The deep neural network DNN employed by the detection appa ^¬ ratus 1 can comprise in a possible embodiment a convolutional neutral network CNN. In an alternative embodiment, the detec- tion apparatus can also employ a recurrent neural network

RNN. In a further alternative embodiment the deep neural net ^¬ work DNN can also comprise a deep fed forward network.

The employed deep neural network DNN is composed of several layers. Each layer can comprise several nodes where computa ^¬ tion is performed. A node of the layer combines input from the data with a set of coefficients or weights that either amplify or dampen the output, thereby assigning significance to inputs for the task the algorithm is trying to learn. The- se products can be summed and the generated sum is passed through a node's so-called activation function to determine whether and to what extent said signal progresses further through the neural network NN to affect the ultimate output. Accordingly, a layer of the employed deep neural network DNN can comprise a row of neuron-like switches that turn on and off as input data is fed through the deep neural network DNN. Each layer' s output is simultaneously the subsequent layer' s input starting from an initial input layer receiving the da- ta. The deep neural network DNN comprises at least one input layer, at least one output layer and one or several hidden layers in-between the input layer and the output layer. In deep learning neural networks, each layer of nodes trains on a distinct set of features based on the previous layer' s output. The further one advances into the deep neural network DNN, the more complex the features become which the nodes can recognize because the nodes aggregate and recombine features from the previous layer. This is also known as feature hier ^¬ archy. The hierarchy is a hierarchy of increasing complexity and abstraction. A feature hierarchy makes deep learning neural networks capable of handling very large, high-dimensional datasets with a plurality of parameters that can pass through non-linear functions.

The deep learning neural network DNN trained on labelled data can after training be applied to unstructured data comprising the digital input image data.

In a possible embodiment, the deep neural network DNN is formed by a convolutional neural network CNN. The convolu- tional neural network CNN is a type of feed forward artifi- cial neural network. A convolutional layer forms a core building block of a convolutional neural network CNN. The layers' parameters consist of a set of learnable filters or kernels which have a small receptive field but extend through the full depth of the input data volume. During the forward pass, each filter is convolved across the width and height of the input volume, computing a dot product between the entries of the filter and the input and producing a two-dimensional activation map of the filter. As a result, the network learns filters that can be activated and it detects some specific type of feature at some spatial position in the input data. A possible structure of the convolutional neural network CNN is also illustrated in context with Fig. 4.

In a possible embodiment, the deep neural network DNN for counting objects can be used by the detection apparatus 1. The deconvolved outputs of learned filters of the trained deep neural network DNN are then used for detection of objects of interest. The objects of interest can for example comprise foreground objects of a specific type such as pedes ^¬ trians or vehicles. In a possible embodiment, to achieve de ^¬ tection of the foreground objects of a specific type a deep neural network DNN used for counting objects can be trained. The filters learned by the different layers of the deep neu ^¬ ral network DNN are then used to analyze an input image for detecting objects of interest such as foreground objects. Though the deep neural network DNN has not been explicitly trained for detecting objects, in order to achieve the task of counting objects, it learns filters that are activated by the foreground objects. This fact can be exploited to achieve the detection of the objects. In order to extract the parts of the input image causing activations in the filters of the deep neural network DNN, deconvolution is used in which suc- cessive unpooling, rectification and filtering can be performed to reconstruct the activity in the layer beneath that layer which gave rise to the chosen activation.

Lower layers of the deep neural network DNN in general rep- resent low-level features. Using these does require addition ^¬ al intelligence to separate the foreground from the back ^¬ ground. With an increasing hierarchy of the layers, higher level concepts are learned until the highest layer of the deep neural network DNN is reached. The highest layer of the deep neural network DNN is most task-specific. A combination of channels from mid-level layers and higher-level layers can be used in a possible embodiment to detect foreground objects in the supplied digital images. Different channels of the mid-level layers and/or high level layers can be activated by different parts of the foreground objects. By combining or merging the deconvolved outputs of the learned filters, it is possible to obtain an indication of specific objects of in ^¬ terest such as foreground objects. Fig. 2 shows a flowchart of a possible exemplary embodiment of a method for detecting objects of interest in images ac ^¬ cording to an aspect of the present invention. The method comprises in the illustrated exemplary embodiment two main steps .

In a first step SI, at least one input image is supplied to a trained deep neural network DNN which comprises a stack of layers .

In a further step S2, the deconvolved output of a learned filter is used or the deconvolved outputs of learned filters of at least one layer of the trained deep neural network DNN are combined in step S2 to detect the object of interest in the supplied images.

Before supplying the input images to the deep neural network DNN, the deep neural network DNN is trained in a training phase. The deep neural network DNN can be trained in a possi ^¬ ble embodiment on annotated training images comprising anno ^¬ tation labels. These annotation labels can include count val ^¬ ues indicating a number of objects of interest within each of the supplied training image. The training of the deep neural network DNN can be performed on the basis of annotated syn ^¬ thetic training images and/or annotated natural training im ^¬ ages . In a possible embodiment, the deconvolved outputs of learned filters of mid-level and high-level layers of the trained deep neural network DNN are combined in step S2 to detect the objects of interest in the supplied digital images. The supplied digital images can comprise pixels with pixel values for different channels, in particular color channels including pixel values for different colors including a red color, a green color, and a blue color. Fig. 3 shows a schematic diagram for illustrating a possible embodiment of a method for detecting objects of interest in images. The illustrated embodiment can for instance be used for detecting pedestrians as foreground objects in supplied digital images. As can be seen in the embodiment of Fig. 3, in initial step SO, a provided deep neural network DNN is trained using training images TIMG supplied to the deep neu ^¬ ral network DNN. The used deep neural network DNN can be for instance a deep neural network DNN for counting objects. Fur ^¬ ther, at least one image such as a test image TE-TMG can be supplied to the trained deep neural network DNN in step SI as illustrated also in Fig. 3. In a possible embodiment, the se ^¬ quence of input images comprising a plurality of input images or input image frames can be supplied to the trained deep neural network DNN. In a further step S2, the deconvolved outputs of learned filters of at least one layer of the trained deep neural network DNN are combined to detect the object of interest in the supplied images. In a possible em- bodiment, deconvolved outputs of learned filters of the trained deep neural network DNN or data model are merged for detection in step S2.

The processing can be extended as illustrated in Fig. 3. In a further step S3, a low-level processing noise reduction and a pseudo-coloring can be performed. The noise of the combined deconvolved outputs of the learned filters can be reduced by processing different channels separately. It is possible that at least one channel is preserved while at least another channel is suppressed to reduce a noise within the combined deconvolved outputs of the learned filters. In a possible em ^¬ bodiment, a noise reduction can be achieved on the composite image by using techniques where different color channels are treated separately. In a possible implementation, the details of red color channels and blue color channels are preserved while suppressing the details of a green color channel since the foreground objects are highlighted in color tones with high values of the red and blue channels. This is because the deconvolution results in features of the input image causing activations to be highlighted in colors dominated by red and blue . The dominant color channels, i.e. the red color channel and the blue color channel, can be the same for all kinds of im ^¬ age date since a heatmap generated after deconvolution for any image would result in the features of the input image causing activations to be highlighted in these colors. Howev ^¬ er, if the process of plotting the heatmap is changed, then the representation can change and the processing can be adapted accordingly. The noise reduction is followed in step S3 by performing a so-called pseudo-coloring. In a possible embodiment, color variations are produced based on sine wave generation for each of the color channels. In a possible implementation, the mean value of the red color channel and the blue color channel are increased to bring out the foreground objects more clearly. Accordingly, selected channels can be processed in a possible embodiment to increase a mean value of the respective channel.

In a further processing step, a background elimination can be performed. In this step S4, a background within the input im- age is eliminated at least partially by comparing the output of a learned filter of a layer with the deconvolved output of a learned filter or a combination of the deconvolved outputs of learned filters For instance, the output from a channel of the convolutional layer 5 of the deep neural network DNN can be used to remove parts of the background.

With the increasing hierarchy of the layers of the deep neu ^¬ ral network DNN, the filters learn better the concept of the foreground. In a possible embodiment, filters in the convolu- tional layer 5 can comprise an output that is able to local ^¬ ize the foreground well while being able to distinguish it from large parts of the background. It is possible to obtain information about the separation from the foreground and background of the image frame at a global level. By comparing the channel output of a specific layer such as the fifth con ^¬ volutional layer of the DNN with the output obtained after low-level processing (noise removal and pseudo-coloring) , parts of the background are eliminated. The employed concept is similar to said of guided filtering where the content of a guide image is used to filter the actual image. Likewise, in this case, the filter output of the specific convolutional layer, such as the fifth convolutional layer of the DNN can serve as the guide image. For regions with low or zero pixel values in the guide image, the output of the image being fil ^¬ tered can be discarded or considered part of the background while the rest is considered as foreground. p(x,y) denotes the pseudo-colored output where p is the pixel value at co ^¬ ordinates x, y in the digital image. g(x,y) denotes the guide image (for example, the output of a filter or channel of a specific convolutional layer such as the fifth convolutional layer where g is the pixel value at co-ordinates x, y of the digital image. In a possible embodiment, p(x,y) is used as a first input and g(x,y), e.g. the guide image, is used as a second input for step S4. The output of step S4 can be denot ^¬ ed by f (x,y) where f is the pixel value at co-ordinates x, y in the digital image, wherein f (x,y) = the background, if g(x,y) ≤ a threshold value and f(x,y) = p(x,y), if g(x,y) > the threshold value.

The background elimination in step S4 is followed in the il ^¬ lustrated embodiment of Fig. 3 by a thresholding step S5. The thresholding is performed to detect contours of objects of interest within the supplied images. If the input of the thresholding step is denoted by f (x,y) wherein f is the pixel value at co-ordinates x, y in the digital image and o(x,y) is the output of the thresholding step wherein o is the pixel value at co-ordinates x, y, in the digital image: o(x,y) = foreground if f (x,y) ≤ a threshold value and

o(x,y) = background if f (x,y) > the threshold value.

A contour detection in step S5 is followed by bounding the contour results in bounding boxes. The bounding is performed in step S6 as illustrated in Fig. 3. In step S6, the bounding boxes are generated at the detected contours of the objects of interest within the supplied images. The bounding boxes can be superimposed on the input image around the detected contours to mark objects of interest within the input images. The input images with the superimposed bounding boxes can be output as illustrated in Fig. 3. The input images with the superimposed bounding boxes can be stored in a memory or transmitted to another device or displayed on a screen to a user . Fig. 3 shows a possible embodiment with an extended

postprocessing. This postprocessing includes low-level processing including noise reduction and pseudo-coloring (step S3) to highlight a foreground region. By using channels from a mid-level layer of the deep neural network DNN, there are also some parts of the background that can cause activations that need to be eliminated. The background elimination is performed in step S4. By using some of the filters learned by the higher layers, parts of the background can be eliminated in step S4. This is followed by thresholding in step S5 to segment the foreground region. Finally, the segmented fore ^¬ ground is analyzed for closed contours and can be enclosed by generated bounding boxes in step S6.

The used deep neural network DNN is trained by using training images TIMG as illustrated in Fig. 3. The deep neural network DNN for counting objects can be trained using synthetic imag ^¬ es and/or natural images. A model can be tuned for the target dataset using few digital images. This avoids the need for large annotated training data from the target site during training.

Transfer learning can be used to train a convolutional neural network CNN for counting pedestrians using synthetic images from a baseline network trained for image classification. There is also the possibility of using natural images or a combination of synthetic images and natural images to train and tune the deep neural network DNN. Fig. 4 shows the structure of a deep neural network DNN which can be used for pedestrian counting. The deep neural network DNN comprises an input data layer IDL as illustrated in

Fig. 4. Further, the deep neural network DNN comprises sever- al convolutional layers convl to conv5 as shown in Fig. 4. The convolutional layers can be followed by fully connected layers fc6, fc7, fc8 as shown in Fig. 4. In the illustrated embodiment, the numbering of the layers is in increasing or ^¬ der of hierarchy with layer 1 being the lowest layer and lay- er 8 indicating the highest layer within the hierarchy. The illustrated blocks and corresponding numbers denote possible dimensions of the volume of neurons or nodes in each layer of the deep neural network DNN. For the five convolutional lay ^¬ ers convl to conv5, a kernel size is also illustrated. For instance, the first convolutional layer 1 of the deep neural network DNN has volume dimensions 55x55x96. The kernel of the first convolutional layer has a kernel size 5x5. The dimen ^¬ sion of the second convolutional layer 2 is 27x27x256 includ ^¬ ing a kernel with a kernel size of 3x3 as shown in Fig. 4. The next convolutional layer conv3 has the dimensions

13x13x384 and a kernel with a kernel size 3x3. The fourth convolutional layer conv4 comprises dimensions 13x13x384 with a kernel size 3x3. The last convolutional layer conv5 com ^¬ prises dimensions 13x13x256 indicating the volume of neurons or nodes in this layer. The input data layer IDL on the left side of Fig. 4 comprises a data layer which is fed to the deep neural network DNN. A convolutional deep neural network CNN can comprise as illustrated in Fig. 4 five convolutional layers and three fully connected layers fc6, fc7, fc8. The final fully connected layer fc8 forms a classifier that gives a probability for each class. Rectified linear units can be used for the activation functions. Pooling and local response normalization layers can be present at the convolutional lay ^¬ ers. A dropout can be used to reduce overfitting.

A forward pass through the deep neural network DNN comprises a data flowing from the input data layer IDL to the output layer, i.e. from the layers lower in the hierarchy to the layers higher in the hierarchy. A deconvolution pass involves starting from one of the learned filters at a layer and doing the reverse process of successive unpooling, rectification and filtering to reconstruct the activity in the layer be- neath that which gives rise to the chosen activation. This flow is from the layers higher in the hierarchy to the layers lower in the hierarchy.

In a different embodiment, the number and dimensions of con- volutional layers and fully connected layers can vary depend ^¬ ing on the use case. The apparatus for detecting objects of interest in digital images can be used for different use cas ^¬ es including driver assistance systems or surveillance sys ^¬ tems. The method for detecting objects of interest in images can be performed in parallel for detecting different kind of objects of interest in images simultaneously. The images are in a preferred embodiment digital colored images which may be captured by cameras of a vehicle. The images captured by the cameras can be supplied to a trained deep neural network DNN as illustrated in the exemplary embodiment of Fig. 4. The deconvolved outputs of learned filters of at least one layer of the trained deep neural network DNN are combined and pro ^¬ cessed by a processing unit to detect objects of interest in the supplied input images. These objects can for instance comprise pedestrians in the surrounding of the vehicle or other vehicles in the surrounding of the vehicle. The image can undergo a postprocessing as illustrated in Fig. 3 to pro ^¬ vide bounding boxes or other symbols to mark objects of in ^¬ terest within the supplied input image. The detected objects of interest such as pedestrians and/or vehicles can be dis ^¬ played with different symbols or bounding boxes on a screen to a user, for instance to a driver of a vehicle. In a possi ^¬ ble embodiment, the employed deep neural network DNN or data model can be loaded from a database for different use cases or operation modes of the system.