Method for reducing video detection errors and host

Abstract

The application provides a processing method and a host computer for reducing image detection errors. The processing method comprises: using an object detection model to classify a range of interest included in an image, so as to find a first-class object pixel range and a second-class object pixel range; comparing the first-class object pixel range and a plurality of first example masks corresponding to the first-class object, and comparing the second-class object pixel range and a plurality of second example masks corresponding to the second-class object, to obtain a first judgment result and a second judgment result; and judging whether the range of interest includes the second-class object according to the first judgment result and the second judgment result. The processing method and the host computer provided by the application can reduce image detection errors, improve the detection of small image proportion objects, reduce false alarm probability, and enhance the reliability of image detection and / or driver monitoring systems.

Description

Technical Field

[0001] This application relates to driving safety, particularly to video monitoring mechanisms related to driver attention levels, specifically to a processing method and host computer for reducing video detection errors. Background Technology

[0002] Recent technological advancements in assisted or autonomous vehicle driving have yielded significant progress. Improved yield rates for on-chip systems have enabled substantial advancements in vehicle cognition, sensing, and control systems, increasing the possibility of more reliable and safer fully automated driving vehicles. However, the presence of a driver remains essential before these ideals can be realized.

[0003] One crucial factor in ensuring vehicle safety is the driver's level of attention, encompassing both cognitive and visual aspects. An analysis of accident reports from Europe, the US, and Australia between 2019 and 2021 revealed that approximately 17% of fatal vehicle accidents were due to driver distraction and / or inattention. According to the US Centers for Disease Control and Prevention (CDC), driver distraction can be categorized into three types: visual distraction (eyesight deviation events), operational distraction (hands deviation events), and psychological distraction (focus deviation events). Examples of driver inattention include interacting with passengers, using the phone, looking inside or outside the vehicle, eating, drinking, singing or dancing, applying makeup, smoking, operating devices, and retrieving items. These behaviors lead to slower reaction times, hindering appropriate planning and control decisions, highlighting the need for a Driver Monitoring System (DMS). Based on events detected by sensors, driver monitoring systems are used to generate feedback for drivers, including auditory, visual, and / or tactile warnings, to notify them when their level of attention falls below a certain range.

[0004] Traditional driver monitoring systems often target mobile phones and cigarettes for image detection. However, mobile phones and cigarettes are small in size and occupy a small percentage of pixels in the overall image. Therefore, directly detecting mobile phones and cigarettes using image detection often fails to identify them.

[0005] Another traditional driver monitoring system detects hands holding a phone or cigarette during image processing. While this increases the detection rate, hand gestures like resting the driver's face on their hand or scratching an itch often resemble holding a phone or cigarette, frequently resulting in false alarms. When a false alarm occurs, the driver monitoring system may use feedback mechanisms such as shaking the seat, vibrating the seatbelt, or issuing audible or visual cues to alert the driver. The system may also intervene in vehicle control, such as slowing down or flashing lights. Compared to not detecting a phone or cigarette, these false alarms can be more uncomfortable for the driver and negatively impact their mood.

[0006] Therefore, there is an urgent need for a processing method that can reduce image judgment errors during image detection, which can both improve the detection of objects with small image proportions and reduce the probability of false alarms, so as to enhance the reliability of driver monitoring systems. Summary of the Invention

[0007] According to one embodiment of this application, a processing method for reducing image detection errors is provided, comprising: classifying a region of interest contained in an image using an object detection model to find pixel ranges of a first type of object and pixel ranges of a second type of object; comparing the pixel range of the first type of object with a plurality of first instance masks corresponding to the first type of object to obtain a first judgment result, which indicates whether the pixel range of the first type of object corresponds to the first type of object; comparing the pixel range of the second type of object with a plurality of second instance masks corresponding to the second type of object to obtain a second judgment result, which indicates whether the pixel range of the second type of object corresponds to the second type of object; and determining whether the region of interest contains the second type of object based on the first judgment result and the second judgment result.

[0008] Preferably, in order to obtain the region of interest, the method further includes: receiving the image; and using the object detection model to detect the first type of object to obtain the region of interest.

[0009] Preferably, in order to obtain the first judgment result, the step of comparing the pixel range of the first type of object with the multiple first example masks corresponding to the first type of object further includes: calculating a first similarity between the shape text of the pixel range of the first type of object and the shape text of the multiple first example masks respectively; and when at least one of the first similarities is greater than a first threshold, the first judgment result is true, otherwise the first judgment result is false.

[0010] Preferably, in order to obtain the second judgment result, the step of comparing the pixel range of the second type of object with the plurality of second example masks corresponding to the second type of object further includes: calculating a second similarity between the shape text of the pixel range of the second type of object and the shape text of the plurality of second example masks respectively; and when at least one of the second similarities is greater than a second threshold, the second judgment result is true, otherwise the second judgment result is false.

[0011] Preferably, in order to obtain the first paradigm mask and the second paradigm mask, the plurality of first paradigm masks and the plurality of second paradigm masks are selected from the training dataset of the object detection model.

[0012] Preferably, in order to obtain better object detection results, the object detection model described above includes a neck network using the GhostPAN architecture.

[0013] Preferably, to avoid the model ignoring small target areas due to severe imbalance between positive and negative samples, which would affect the accuracy of image segmentation, the object detection model described above uses Dice loss as the loss function for training.

[0014] Preferably, to avoid false alarms, when both the first and second judgment results are true, the scope of interest is determined to include the second type of object; and when the first judgment result is true and the second judgment result is false, the scope of interest is determined not to include the second type of object.

[0015] Preferably, in order to increase the probability of detecting the second type of object, when the first judgment result is false and the second judgment result is true, it is determined that the range of interest includes the second type of object.

[0016] Preferably, for use in a driver monitoring system, the image is directed at the front of the driver of the vehicle, wherein the first type of object includes one or any combination of the following objects: hands, feet, and a face other than the driver, and wherein the second type of object includes one or any combination of the following objects: mobile phone; cigarette; water cup; water bottle; food; food packaging; eyebrow pencil; makeup mirror; cotton pad; handkerchief; and toilet paper.

[0017] Preferably, in order to prompt the driver, when it is determined that the area of ​​interest includes the second type of object, one of the following prompts is issued: visual prompt; auditory prompt; tactile prompt; the aforementioned tactile prompt includes vibrating or shaking the driver's steering wheel, seat, seat belt, and control device.

[0018] According to one embodiment of this application, a host computer for reducing image detection errors is provided. The host computer includes at least one processor to execute instructions in non-volatile memory for implementing the image detection error reduction processing method as described above.

[0019] Beneficial effects: The processing method and host for reducing image detection errors provided in this application can reduce image judgment errors, improve the detection of objects with small image ratios, and reduce the probability of false alarms, thereby enhancing the reliability of image detection and / or driver monitoring systems. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a driver monitoring system 100 according to an embodiment of this application.

[0021] Figure 2 This is a block diagram of the driver monitoring system 100.

[0022] Figure 3 This is a flowchart illustrating a processing method 300 for reducing image judgment errors according to an embodiment of this application.

[0023] Figure 4 This is a schematic diagram of the classified pixel ranges of multiple regions of interest according to an embodiment of this application.

[0024] Explanation of reference numerals in the attached drawings: 100, driver monitoring system; 110, main unit; 120, camera device; 130, radio device; 140, display device; 150, speaker device; 199, driver; 300, visual attention assessment method; 310-380, steps. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be described in detail below. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] The terms "first," "second," "third," etc. (if present) in the specification, claims, and drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the objects described herein may be used interchangeably where appropriate. In the description of this application, "plural" means two or more, unless otherwise expressly and specifically defined. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. Some block diagrams shown in the figures are functional entities and do not necessarily correspond to physically or logically independent entities; such functional entities may be implemented in software, in one or more hardware circuits or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0027] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between components; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the aforementioned terms in this application according to the specific circumstances.

[0028] To make the objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the drawings and specific embodiments.

[0029] Please refer to Figure 1 The diagram shown is a schematic representation of a driver monitoring system 100 according to an embodiment of this application. Please refer to... Figure 2 As shown, it is a block diagram of the driver monitoring system 100. Although Figure 1 The driver monitoring system 100 shown is installed inside a car body; however, those skilled in the art will understand that the driver monitoring system 100 can be applied to all vehicles, such as locomotives, tricycles, aircraft, sea vehicles, and spacecraft, as long as there is at least one driver 199 in the vehicle to control the vehicle.

[0030] The aforementioned driver monitoring system 100 includes a main unit 110, and at least one camera 120, at least one radio receiver 130, at least one display device 140, and at least one speaker 150 connected to the main unit 110. Figure 1As shown, the field of view of the photographic device 120 includes at least the face or facial features of the driver 199. The sound pickup range of the microphone 130 includes at least the mouth of the driver 199. The driver 199 can see the visual content displayed on the display device 140 and hear the auditory content emitted by the speaker device 150. In one embodiment, the display device 140 is a head-up display in front of the driver. In one embodiment, the speaker device 150 is located in a headrest behind the driver's head.

[0031] In one embodiment, the display device 140 may include a forward-facing head-up display (HUD). In other embodiments, the display device 140 may include a virtual reality, augmented reality, or mixed reality display. Those skilled in the art will understand that even a simple lighting device can serve as the aforementioned display device 140, as long as it can provide visual cues to the driver 199. Similarly, even a simple buzzer can serve as the aforementioned speaker device 150, as long as it can provide auditory cues to the driver 199.

[0032] In some embodiments, the display device 140 is not entirely controlled by the driver monitoring system 100. For example, the display device 140 may be a standalone display device that receives input from the driver monitoring system 100 and other systems, and is responsible for merging display content from multiple sources. Similarly, in some embodiments, the speaker device 150 is not entirely controlled by the driver monitoring system 100. For example, the speaker device 150 may be a standalone audio device that receives input from the driver monitoring system 100 and other systems, and is responsible for merging audio content from multiple sources for playback.

[0033] In one embodiment, the photographic device 120 may include an infrared sensor. In another embodiment, the photographic device 120 may also include one or more infrared lights to allow the infrared sensor to clearly capture the driver 199's face and features in low-visibility conditions. In one embodiment, the microphone 130 may include directional microphone functionality to be aimed at the driver 199's face, particularly the mouth, to filter out external noise. For example, the microphone 130 may include an array of multiple microphone components capable of focusing or scanning the direction of sound wave origin using signal processing. One or more of the microphones 130 may also be used to collect background noise to determine whether the sound originates from the background or the driver 199.

[0034] The aforementioned host 110 may include one or more processors 111 to control the entire host 110 and the driver monitoring system 100. The processor 111 may execute an operating system, as well as drivers and applications running within that operating system environment. This software and / or hardware may be stored in non-volatile memory, containing instructions and data that the processor 111 can execute, for implementing the methods and functions provided in this application. Those skilled in the art should possess the common knowledge of computer organization, computer architecture, operating systems, and system software acquired at the undergraduate level, and can understand that various possible variations and derivatives of the host 110's hardware and software are applicable to this application.

[0035] To accelerate the execution of artificial intelligence and / or neural networks for real-time driver monitoring, the host 110 may include one or more co-processors 112 for providing vectorized arithmetic and logical operation capabilities. The aforementioned neural network may include one or more deep neural networks (DNNs) or a collection of deep neural networks. The aforementioned neural network may include convolutional neural networks (CNNs) or their derivatives. The aforementioned co-processor 112 may be a general-purpose graphics processing unit (GPU) or a dedicated neural network processor (NPU). Those skilled in the art will understand that such co-processors for edge AI are already in practical use, and therefore will not be described in detail here.

[0036] The host 110 may also include one or more peripheral device connection devices 113 for connecting the at least one camera device 120, at least one radio device 130, at least one display device 140, and at least one speaker device 150. The peripheral device connection device 113 may include one or more interfaces conforming to one or more industry standards, such as Bluetooth, USB, IEEE 1394, UART, iSCSI, SATA, PCI-Express, PCT, etc.; it may also include proprietary interfaces.

[0037] The driver monitoring system 100 provided in this application uses a visual sensor such as a camera device 120 to assess the driver's visual attentiveness and an auditory sensor such as a microphone device 130 to assess cognitive attentiveness, in order to provide the driver's overall level of attentiveness. By fusing information from multiple modalities such as vision and hearing, the driver monitoring system 100 provided in this application can infer the driver's visual and cognitive attentiveness.

[0038] Unlike traditional driver monitoring systems that only include optical sensors, the driver monitoring system 100 provided in this application can provide more reliable and valuable information on driver concentration by analyzing the driver's cognitive attention. The driver monitoring system 100 can estimate the driver's visual attention using an image containing the driver's face. This estimation can be performed using an algorithm to determine the driver's face and eye direction from the image. To estimate the driver's cognitive attention, an algorithm can also be used to estimate auditory events based on sound. These auditory events can be categorized into various types. For example, the algorithm can determine whether the driver is in a noisy or distracting environment. Finally, when it is determined that the driver's attention is low based on visual and cognitive attention, the driver can be alerted or warned via the display device 140 and / or the speaker device 150.

[0039] Please refer to Figure 3 The diagram shown is a flowchart illustrating a processing method 300 for reducing image judgment errors according to an embodiment of this application. This processing method 300 for reducing image judgment errors can be implemented by... Figure 1 and Figure 2 The driver monitoring system 100 shown is implemented, particularly by a program executed by the host 110, such as the program executed by the aforementioned processor 111 and coprocessor 112. This application does not restrict the order of execution of any two steps if there is no causal relationship between them. The processing method 300 for reducing image judgment errors may begin at step 310.

[0040] Step 310: Receive image. The image may come from the aforementioned photographic device 120.

[0041] Step 320: Detect a first type of object in the image using an object detection model. The first type of object may include hands, feet, and profile views. When detecting the first type of object, a second type of object that might affect the driver's concentration may also be detected, such as mobile phones, cigarettes, water cups, water bottles, food, food packaging, eyebrow pencils, makeup mirrors, cotton pads, handkerchiefs, and toilet paper.

[0042] Those skilled in the art will understand that the object detection model described above can use common object detection models, such as various versions of the YOLO (You Only Look Once) model, or various convolutional neural network (CNN) models. In one embodiment, the object detection model described above can be a variation of the YOLO model. For example, the GhostPAN architecture is used in the neck network of the YOLO model. The GhostPAN architecture described above is mainly based on the concepts of Ghost modules and Path Aggregation Networks (PANs). The Ghost module typically obtains basic features first through a portion of standard convolutional networks or depth-wise convolutional networks. Then, it reduces computational cost with cheaper linear transformations, such as 1×1 convolutions or other lightweight operations. The main purpose of the Path Aggregation Network (PAN) is to perform bidirectional fusion of multi-layer features generated by the backbone network of the YOLO model in an object detection or segmentation system, that is, to include top-down and bottom-up fusion, in order to enhance the information flow between features at different levels.

[0043] In one embodiment, the object detection model described above can use Dice loss as the loss function. Dice loss was proposed by Fausto Milletari et al. in *V-net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation*, and it originates from the Sorensen-Dice coefficient parameter. Common loss functions used in deep network training for image segmentation often employ cross-entropy or its variations as the primary loss function. However, when the dataset is severely imbalanced between positive and negative samples, relying solely on cross-entropy as the loss function can cause the model to ignore small target regions, thus affecting the accuracy of image segmentation. To address this issue, DSC (Dice Similarity Coefficient) can be used, which is a metric for evaluating the similarity between two samples, ranging from 0 to 1, with larger values ​​indicating greater similarity. The Dice loss function is simply 1 minus the Dice similarity coefficient (DSC).

[0044] According to an embodiment of this application, the method for learning an object detection model using Dice loss is as follows: First, the images and their annotations from the training set are input into the object detection model to initialize its network parameters. Then, the predicted probabilities and ground truth annotations are obtained. This step includes obtaining the target probability p corresponding to each pixel through forward propagation. i And paired with real label y i Then, calculate the aforementioned Dice similarity coefficient (DSC). For the binary segmentation scenario, the Dice similarity coefficient (DSC) can be calculated as follows:

[0045]

[0046] Where pi and yi represent the predicted and actual pixel values ​​(or probabilities) respectively, N is the total number of pixels, and ε is a smoothing term to avoid a denominator of zero. Using the Dice similarity coefficient DSC calculated above, the Dice loss can be obtained as follows:

[0047] Dice loss = 1 - DSC or

[0048]

[0049] The training methods described above can more efficiently learn accurate segmentation of small or unbalanced target regions during the training process, significantly improving the accuracy and reliability of the target detection model.

[0050] If a first-type object is detected in step 320, the process can proceed to step 330. If no first-type object is detected in step 320, the process can end.

[0051] Step 330: Obtain one or more regions of interest in the image based on the detected first type of object. The aforementioned regions of interest are usually rectangular areas, which are a portion of the acquired image.

[0052] Step 340: Input each region of interest into the object detection model described above, and use the object detection model to classify each pixel in the region of interest in order to find the pixel ranges of the first type of object and the second type of object.

[0053] According to one embodiment of this application, classification evaluation is performed using either Dice or mean intersection overunion (mIOU). As mentioned earlier, the Dice similarity coefficient (DSC) is primarily used to measure the degree of overlap between two sets (i.e., the predicted pixel range and the true pixel range). The Dice similarity coefficient (DSC) value ranges between 0 and 1, where 1 indicates that the two sets are completely identical, and 0 indicates that the two sets are completely different. In the field of image segmentation, when A is the pixel range of the ground truth set and B is the pixel range of the prediction set, the Dice similarity coefficient (DSC) can be expressed as:

[0054]

[0055] Where |A∩B| represents the intersection of two sets A and B, which is the number of pixels that are simultaneously classified as belonging to that category, while |A| and |B| represent the number of pixels in sets A and B, respectively.

[0056] The average crosslinking ratio (mIOU) is a commonly used quantitative metric for multi-class segmentation problems. It involves first calculating the crosslinking ratio for each class, then averaging the crosslinking ratios of all classes. When the segmentation problem involves only binary segmentation, the crosslinking ratio for that class is the average crosslinking ratio. For multi-class segmentation, the average crosslinking ratio can simultaneously evaluate the segmentation quality of each class. If we consider A... c B represents the pixel range of category c in the real set A. c Let represent the pixel range of category c in prediction set B. Then, the cross-linking ratio (IOU) for category c can be expressed as:

[0057]

[0058] Where |A c ∪B c | represents the number of pixels in the intersection of the true set and the predicted set for category c, |A c ∪B c | represents the number of pixels in the union of the true set and the predicted set for category c.

[0059] If there are C categories (with or without background), then the overall average crosslinking ratio mIOU can be defined as follows:

[0060]

[0061] Please refer to Figure 4 As shown, it is a schematic diagram of the classified pixel ranges of multiple regions of interest according to an embodiment of this application. Figure 4In this embodiment, objects can be divided into a first category (i.e., a hand), a second category (i.e., cigarettes and a mobile phone), and a background. Pixels classified as mobile phones are blue, and pixels classified as cigarettes are white. Pixels classified as hands are pink. The remaining pixels belong to the background.

[0062] According to an embodiment provided in this application, the aforementioned object detection model can be evaluated using the Dice and the average crosslinking ratio mIOU, as shown in Table 1.

[0063] Table 1. Classification accuracy of the object detection model

[0064] Average crosslinking ratio mIOU 90.2％

[0065] Step 350: Compare the pixel range of the first type of object with the multiple first instance masks corresponding to the first type of object to determine whether the pixel range of the first type of object corresponds to the first type of object, and obtain the first judgment result.

[0066] The aforementioned multiple first-example masks all correspond to the aforementioned first-class objects. That is, within the test dataset of the object detection model, the similarity of features of the first-class objects is calculated by comparison, and multiple representative first-example masks are selected. In one embodiment, the aforementioned similarity calculation method can employ ShapeContext, a classic method for two-dimensional shape comparison or contour matching, which can also be applied to other scenarios involving three-dimensional point clouds.

[0067] The first step in shape text extraction is to extract contour points, which are two-dimensional masking or boundary points of the object extracted from the image and sampled into several discrete points. The second step is to construct a local shape description. For each discrete point, the distance and direction distribution relative to other points are statistically analyzed, forming a statistical histogram separated by logarithmic radius and angle, which is called shape text.

[0068] When comparing whether two shapes match, one can first obtain the shape text of both shapes. Then, using methods such as bipartite matching, one can find the point correspondences that minimize the overall matching error. The accumulated total error of the matching results, which is a weighted sum of the distances between corresponding points, can be used to determine the degree of similarity between the two shapes. Those skilled in the art will understand that comparing the similarity of two shapes using shape text is a conventional method.

[0069] In this step, a first threshold can be set. When the similarity between the shape of the pixel range of the first type of object and the shape of one of the multiple first example masks exceeds the first threshold, it can be determined that the pixel range of the first type of object corresponds to the first type of object, and the first determination result is true. If the similarity between the shape of the pixel range of the first type of object and the shapes of all the multiple first example masks does not exceed the first threshold, it can be determined that the pixel range of the first type of object does not correspond to the first type of object, and the first determination result is false.

[0070] Step 360: Compare the pixel range of the second type of object with multiple second-example masks corresponding to the second type of object to determine whether the pixel range of the second type of object corresponds to the second type of object, and obtain a second judgment result. Similarly, the multiple second-example masks mentioned above all correspond to the second type of object mentioned above. That is, the similarity of the features of the second type of object is calculated by comparing them within the test dataset of the object detection model, and multiple representative second-example masks are selected.

[0071] In this step, a second threshold can be set. When the similarity between the shape of the pixel range of the second type of object and the shape of one of the multiple second-example masks exceeds the second threshold, it can be determined that the pixel range of the second type of object corresponds to the second type of object, and the second determination result is true. If the similarity between the shape of the pixel range of the second type of object and the shapes of all the multiple second-example masks does not exceed the second threshold, it can be determined that the pixel range of the second type of object does not correspond to the second type of object, and the second determination result is false.

[0072] Step 370: Based on the first and second judgment results described above, determine whether the scope of interest actually includes the second type of object. When both the first and second judgment results are true, it can be determined that the scope of interest includes the second type of object. When the first judgment result is true, but the second judgment result is false, it can be determined that the scope of interest does not include the second type of object. This can reduce false alarms.

[0073] When the first judgment result is false, but the second judgment result is true, that is, only second-type objects such as mobile phones or cigarettes are detected, but first-type objects such as hands are not detected. This might be because the second-type object is placed somewhere, but the driver did not use their hand to pick it up. Or it might be because the vehicle's movement caused the second-type object to fly up. In some embodiments, even if the second-type object is detected, it can be determined that the area of ​​interest does not include the second-type object. Those skilled in the art will understand that when the first judgment result is false, but the second judgment result is true, it can also be determined that the area of ​​interest includes the second-type object, in order to increase the probability of detecting the second-type object.

[0074] Step 380: When it is determined that the scope of interest includes the second type of object, a notification is issued. Those skilled in the art will understand that the notification may include visual cues using a display device 140 or lights, auditory cues using a speaker 150, or notifications to external systems, such as vehicle driving systems, via an electronic interface.

[0075] According to one embodiment of this application, a processing method for reducing image detection errors is provided, comprising: classifying a region of interest contained in an image using an object detection model to find pixel ranges of a first type of object and pixel ranges of a second type of object; comparing the pixel range of the first type of object with a plurality of first instance masks corresponding to the first type of object to obtain a first judgment result, which indicates whether the pixel range of the first type of object corresponds to the first type of object; comparing the pixel range of the second type of object with a plurality of second instance masks corresponding to the second type of object to obtain a second judgment result, which indicates whether the pixel range of the second type of object corresponds to the second type of object; and determining whether the region of interest contains the second type of object based on the first judgment result and the second judgment result.

[0076] Preferably, in order to obtain the region of interest, the method further includes: receiving the image; and using the object detection model to detect the first type of object to obtain the region of interest.

[0077] Preferably, in order to obtain the first judgment result, the step of comparing the pixel range of the first type of object with the multiple first example masks corresponding to the first type of object further includes: calculating a first similarity between the shape text of the pixel range of the first type of object and the shape text of the multiple first example masks respectively; and when at least one of the first similarities is greater than a first threshold, the first judgment result is true, otherwise the first judgment result is false.

[0078] Preferably, in order to obtain the second judgment result, the step of comparing the pixel range of the second type of object with the plurality of second example masks corresponding to the second type of object further includes: calculating a second similarity between the shape text of the pixel range of the second type of object and the shape text of the plurality of second example masks respectively; and when at least one of the second similarities is greater than a second threshold, the second judgment result is true, otherwise the second judgment result is false.

[0079] Preferably, in order to obtain the first paradigm mask and the second paradigm mask, the plurality of first paradigm masks and the plurality of second paradigm masks are selected from the training dataset of the object detection model.

[0080] Preferably, in order to obtain better object detection results, the object detection model described above includes a neck network using the GhostPAN architecture.

[0081] Preferably, to avoid the model ignoring small target areas due to severe imbalance between positive and negative samples, which would affect the accuracy of image segmentation, the object detection model described above uses Dice loss as the loss function for training.

[0082] Preferably, to avoid false alarms, when both the first and second judgment results are true, the scope of interest is determined to include the second type of object; and when the first judgment result is true and the second judgment result is false, the scope of interest is determined not to include the second type of object.

[0083] Preferably, in order to increase the probability of detecting the second type of object, when the first judgment result is false and the second judgment result is true, it is determined that the range of interest includes the second type of object.

[0084] Preferably, for use in a driver monitoring system, the image is directed at the front of the driver of the vehicle, wherein the first type of object includes one or any combination of the following objects: hands, feet, and a face other than the driver, and wherein the second type of object includes one or any combination of the following objects: mobile phone; cigarette; water cup; water bottle; food; food packaging; eyebrow pencil; makeup mirror; cotton pad; handkerchief; and toilet paper.

[0085] Preferably, in order to prompt the driver, when it is determined that the area of ​​interest includes the second type of object, one of the following prompts is issued: a visual prompt; an auditory prompt; a tactile prompt. The aforementioned tactile prompt includes vibrating or shaking the driver's steering wheel, seat, seat belt, and control devices.

[0086] According to one embodiment of this application, a host computer for reducing image detection errors is provided. The host computer includes at least one processor to execute instructions in non-volatile memory for implementing the image detection error reduction processing method as described above.

[0087] The processing method and host for reducing image detection errors provided in this application can reduce image judgment errors, improve the detection of objects with small image ratios, and reduce the probability of false alarms, thereby enhancing the reliability of image detection and / or driver monitoring systems.

Claims

1. A processing method for reducing image detection errors, characterized in that, Include: The object detection model is used to classify the regions of interest contained in an image in order to find the pixel ranges of the first type of object and the pixel ranges of the second type of object; The pixel range of the first type of object is compared with multiple first instance masks corresponding to the first type of object to obtain a first judgment result, which is used to indicate whether the pixel range of the first type of object corresponds to the first type of object; The second judgment result is obtained by comparing the pixel range of the second type of object with multiple second instance masks corresponding to the second type of object, which indicates whether the pixel range of the second type of object corresponds to the second type of object; as well as Based on the first judgment result and the second judgment result, determine whether the scope of interest includes the second type of object.

2. The image detection error reduction processing method as described in claim 1, characterized in that, It further includes: receiving the image; and using the object detection model to detect the first type of object to obtain the range of interest.

3. The image detection error reduction processing method as described in claim 1, characterized in that, The aforementioned step of comparing the pixel range of the first type of object with the multiple first example masks corresponding to the first type of object further includes: calculating a first similarity using the shape text of the pixel range of the first type of object and the shape text of the multiple first example masks respectively; and when at least one of the first similarities is greater than a first threshold, the first judgment result is true, otherwise the first judgment result is false. The above-mentioned step of comparing the pixel range of the second type of object with the multiple second example masks corresponding to the second type of object further includes: calculating a second similarity between the shape text of the pixel range of the second type of object and the shape text of the multiple second example masks respectively; and when at least one of the second similarities is greater than a second threshold, the second judgment result is true, otherwise the second judgment result is false.

4. The processing method for reducing image detection errors as described in claim 1, characterized in that, The aforementioned first-paradigm masks and the aforementioned second-paradigm masks are selected from the training dataset of the aforementioned object detection model.

5. The processing method for reducing image detection errors as described in claim 1, characterized in that, The object detection model described above includes a neck network using the GhostPAN architecture.

6. The processing method for reducing image detection errors as described in claim 1, characterized in that, The object detection model described above is trained using Dice loss as the loss function.

7. The processing method for reducing image detection errors as described in claim 1, characterized in that, When both the first and second judgment results are true, it is determined that the scope of interest includes the second type of object; and when the first judgment result is true and the second judgment result is false, it is determined that the scope of interest does not include the second type of object.

8. The processing method for reducing image detection errors as described in claim 1, characterized in that, When the first judgment result is false and the second judgment result is true, it is determined that the scope of interest includes the second type of object.

9. The processing method for reducing image detection errors as described in claim 1, characterized in that, The image is a frontal view of the vehicle's driver, wherein the first category of objects includes one or any combination of the following objects: hands, feet, and a face other than the driver, and the second category of objects includes one or any combination of the following objects: a mobile phone; a cigarette; a water cup; a water bottle; food; food packaging; an eyebrow pencil; a makeup mirror; cotton pads; a handkerchief; and toilet paper.

10. The processing method for reducing image detection errors as described in claim 1, characterized in that, When it is determined that the area of ​​interest includes the second type of object, one of the following prompts is issued: visual prompt; auditory prompt; tactile prompt; the aforementioned tactile prompt includes vibration or shaking of the driver's steering wheel, seat, seat belt, and control device.

11. A host for reducing image detection errors, the host comprising at least one processor to execute instructions in non-volatile memory for implementing the processing method for reducing image detection errors as described in any one of claims 1-10.

Images