Identifying potential false alarm detection boxes

Abstract

Identifying potential false alarm detection boxes is disclosed. The invention relates to an anchor-based object detection system and method that identifies potential false alarm detection among three detection boxes in the same image frame. Each detection box has a predicted IoU score representing its confidence of capturing an object. The system determines (S502) an overlap of a second frame with the first and third frames, with the second frame located therebetween. If it determines (S506) that the IoU score of the second frame is below a set threshold, and if it determines (S508) that the corresponding reference point within each frame is substantially aligned, the reference point of the second frame is close to an alignment line defined by the first frame and the third frame, it identifies (S504) the second frame as a potential false alarm.

Description

Technical Field

[0001] This disclosure relates to object detection, and more specifically, to methods, systems, and software for identifying potential false alarm detection boxes in an anchor-based object detection system. Background Technology

[0002] In modern object detection systems such as Single Sight Detector (SSD) and YOLO (You Only See Once), anchor boxes are fundamental for detecting objects in an image. These anchor boxes are predefined and typically cover the image at various scales and aspect ratios to detect objects of different sizes and shapes. During training, the object detection system learns to adjust these anchor boxes to better fit objects by encoding those with high Intersection over Union (IoU) scores, which represent the overlap between the anchor box and the real data object. For training purposes, anchor boxes with significant IoU overlap are assigned to the object.

[0003] Significant problems can arise when multiple objects are located close to each other, or when anchor boxes are sparsely distributed across an image. In such cases, more than one object may utilize specific anchor boxes and have similar IoUs, leading to ambiguous assignments during training. This ambiguity can result in a phenomenon known as intermediate detection.

[0004] When two or more objects share similar IoU scores using the same anchor box, the object detection system may inconsistently assign anchor boxes to different objects during training. This leads to intermediate detections that are either false positives or ambiguous detections (incorrectly detected boxes located between real objects). These intermediate detections negatively impact the performance of the object detection system by introducing false positives.

[0005] Therefore, there is a need for improvement in this context. Summary of the Invention

[0006] In view of the foregoing, it would be advantageous to resolve or at least mitigate one or more of the aforementioned disadvantages, as set forth in the appended independent patent claims.

[0007] According to a first aspect of the invention, a method is provided for identifying a potentially false positive detection box in a set of three detection boxes within an anchor-based object detection system, wherein each of the three detection boxes is detected in the same image frame, wherein each of the three detection boxes is associated with a corresponding predicted intersection-over-union (IoU) score, wherein the predicted IoU score indicates the confidence of the anchor-based object detection system that the detection box represents an object, the method comprising: determining that a second detection box at least partially overlaps with both a first detection box and a third detection box, wherein the second detection box is located between the first detection box and the third detection box in the image frame; identifying the second detection box as a potentially false positive detection box by: determining that the predicted IoU score of the second detection box is below a first threshold score; and determining a first reference point in the first detection box, a corresponding second reference point in the second detection box, and a corresponding third reference point in the third detection box, and determining that the first reference point, the second reference point, and the third reference point are substantially aligned in the image frame such that the second reference point is within a threshold distance from the alignment defined by the first reference point and the third reference point.

[0008] This disclosure addresses the problem of intermediate detections caused by ambiguous anchor box assignment during the training of anchor-based object detection systems, particularly when anchor boxes are sparsely distributed. The techniques described herein aim to improve detection accuracy by identifying potential false positive detection boxes caused by such ambiguity. Specifically, the method focuses on identifying detection boxes that are likely to fall between real objects.

[0009] This method introduces a strategy for managing fuzzy detection boxes while minimizing computational impact, thereby allowing the object detection system to maintain reliable performance even under hardware constraints or when anchor boxes are sparsely distributed. By identifying these fuzzy detection boxes, the object detection system is better configured to handle false alarms, thus improving accuracy and efficiency.

[0010] In object detection systems like SSD or YOLO, the "predicted IoU score" refers to a metric predicted by the object detection system / model that indicates the degree to which the proposed detection (bounding box) might overlap with a real object in an image. IoU (Intersection over Union) traditionally refers to the ratio of the area of ​​overlap between the predicted bounding box and the ground truth data box to the area of ​​their union. However, in this context, the predicted IoU score acts as a confidence measure of the likelihood that the bounding box generated by the object detection model corresponds to a real object before any post-processing. Nevertheless, IoU prediction can be turned off because the object detection system attempts to predict the geometric IoU between the predicted bounding box and the "imagined" ground truth data bounding box.

[0011] The predicted IoU score is sometimes also called the "object-degree score". Therefore, the object-degree score or predicted IoU score represents the probability or confidence that a given bounding box (detection box) contains an object, rather than just the background.

[0012] To filter out detection boxes and check whether they are potential intermediate detections, the method first verifies that a detection box at least partially overlaps with both the first and third detection boxes, with the second detection box located between the first and third detection boxes in the image frame. This serves as an effective first filtering step before applying more complex analysis.

[0013] Secondly, the predicted IoU score is evaluated. A low predicted IoU score can indicate that the detection box is a potentially ambiguous detection (potential intermediate detection, potential false positive). However, the inventors have recognized that the predicted IoU score provided by the object detection system is not always a reliable indicator. Relying solely on filtering based on low predicted IoU scores (i.e., below a predefined threshold) can lead to a high false negative rate, as actual object detections may be misclassified as false positives.

[0014] The inventors have also recognized that ambiguous detection boxes resulting from inconsistent assignment of anchor boxes to different objects during training are often approximately on the same path as two neighboring detection boxes. Alignment is determined by identifying corresponding reference points within the three detection boxes, meaning that the reference points are in the same relative position in each box (e.g., center, top left corner, or middle of the top edge, etc.). In other words, the term "same relative position" refers to a consistent and predetermined position within each detection box. This can be more precisely expressed as the "same normalized position" representing fixed coordinates within the box (e.g., [0.5, 0.5] relative to the center), which remains constant across different boxes regardless of their size or position within the image frame. Other suitable expressions include "same geometric position," "uniform position anchor," or "normalized reference position." Thus, the reference points are not arbitrarily placed but are consistently defined within each box to enable alignment analysis of the detection boxes.

[0015] The three reference points are then evaluated to determine if they are substantially aligned within the image frame. To evaluate this, the second reference point (which belongs to the potentially ambiguous detection box) is checked to see if it is within a specified threshold distance from the alignment defined by the first and third reference points. This threshold allows for small deviations, ensuring that the system can adapt to minor variations in the positioning of the three detection boxes while still recognizing the overall alignment between them.

[0016] According to some examples, determining that the first reference point, the second reference point, and the third reference point are substantially aligned includes: determining a first vector between a first pair of reference points selected from the first reference point, the second reference point, and the third reference point, and a second vector between different second pairs of reference points selected from the first reference point, the second reference point, and the third reference point, wherein the first reference point, the second reference point, and the third reference point are substantially aligned in the image frame if the absolute value of the cosine of the angle θ between the first vector and the second vector is less than a threshold value of 1.

[0017] Advantageously, the cosine of the angle directly measures the relative orientation of the reference point, thus making the alignment assessment both accurate and consistent.

[0018] Based on some examples, determining that the first reference point, the second reference point, and the third reference point are substantially aligned includes: determining that the second reference point is within a threshold distance from the line formed between the first reference point and the third reference point.

[0019] Advantageously, checking whether a checkpoint is offline within a threshold distance is a low-complexity, orientation-independent geometric operation. It can be implemented in systems with limited computational resources, thus benefiting real-time processing or hardware-constrained environments.

[0020] According to some embodiments, identifying the second detection box as a potential false alarm detection is further performed by the following steps: determining that the predicted IoU score of the second detection box is at least a second threshold score lower than the predicted IoU score of each detection box in the first and third detection boxes.

[0021] Advantageously, using relative comparisons of predicted IoU scores can improve the accuracy of identifying potential false positives. This approach uses the difference in IoU scores between neighboring detection boxes, rather than relying solely on an absolute IoU threshold that could lead to misclassification.

[0022] In some examples, each of the three detection boxes is associated with a predicted object category. Identifying a potential false positive detection in the second detection box is further performed by determining that the predicted object category associated with each of the first, second, and third detection boxes is the same. When two objects of the same category utilize specific anchor boxes and have similar IoUs, the problem of intermediate detections may occur more frequently during the training of the object detection system. Therefore, filtering based on the three detection boxes associated with the same category can provide a more effective means of identifying intermediate detections, leading to more accurate object detection.

[0023] In some examples, the first, second, and third reference points are the midpoints of the top edges of the first, second, and third detection boxes, respectively. For instance, when the camera is capturing an object from the side or front, meaning from a horizontal or angular viewpoint in the image frame rather than from above, the top edges of the detection boxes provide a clear and stable reference for alignment.

[0024] In some examples, the first, second, and third reference points are the center points of the first, second, and third detection boxes, respectively. This approach can be particularly effective when the camera is capturing an object from a top-down perspective. In this case, the center of each detection box provides a clear and stable reference for alignment.

[0025] In some examples, the method further includes assigning a lower probability associated with an object trajectory in the object tracking system to a second detection box compared to the probabilities assigned to the first and third detection boxes.

[0026] In these examples, the method does not necessarily discard the second detection box because there may not be enough certainty that it is a false alarm; it can still represent a valid object. Instead, the second detection box marked as a potential false alarm is deweighted in subsequent post-analysis, such as object tracking. This approach ensures that the second detection box is not prioritized during tracking, but is still considered if it represents an actual object. By reducing the likelihood of the second detection box being associated with the object trajectory, the method can mitigate the risk of associating false alarms with the object trajectory while preventing the premature discarding of potentially valid detections, thereby promoting improved accuracy and robustness in the tracking process.

[0027] In some examples, assigning a lower probability includes assigning a higher cost associated with the object trajectory to the second detection box compared to the cost associated with the first or third detection box.

[0028] In object tracking systems, a cost function is typically used to determine the probability or confidence of matching a detection box with an existing object trajectory. By assigning a higher cost to a second detection box (e.g., adding or multiplying the cost by a constant value), this method effectively reduces the probability of associating it with the trajectory, thus indicating a lower confidence that the detection corresponds to a real object and making it less likely to be selected unless strongly supported by other evidence (feature vector similarity, location, etc.). The matching cost can be further based on feature distance, motion information, etc. The association algorithm can, for example, be greedy, i.e., starting by matching the box with the lowest cost.

[0029] In some examples, assigning a lower probability includes: assigning a second detection box to a lower priority partition of the detection box associated with the object trajectory, and assigning the first and third detection boxes to a higher priority partition of the detection box associated with the object trajectory, wherein the partitions are processed sequentially to be associated with the trajectory in the object tracking system.

[0030] Advantageously, the method only proceeds to process lower-priority partitions containing potential false positive detection boxes, such as second detection boxes, after attempting to associate a trajectory with a higher-priority detection box. As a result, the robustness of the object tracking system can be improved by prioritizing more certain object detections while still considering ambiguous object detections. This method reduces the risk of associating false positives with object trajectories, potentially leading to more accurate tracking results.

[0031] In some examples, the method further includes filtering out second detection boxes that are marked as potential new object trajectories in the object tracking system from the set of detection boxes in the first image frame.

[0032] In object tracking, when a detection box is detected in an image frame that does not match any existing object trajectory (meaning the detection box does not correspond to an object that has already been tracked), these mismatched boxes can be considered as candidates for creating new object trajectories, which represent the objects that will be tracked in subsequent image frames. By filtering out second detection boxes that may be false alarms as described above (where they do not match existing trajectories), this method reduces the likelihood of creating new object trajectories due to erroneous detections. This can lead to a more accurate tracking system that maintains valid trajectories and avoids tracking non-existent objects. Advantageously, the risk of creating incorrect or unnecessary object trajectories can be reduced.

[0033] In some examples, the method further includes counting the first and third detection boxes as confirmed objects in the object counting system, and counting the second detection box as an indeterminate object in the same system. This means that for object counting purposes, the more confidently identified first and third detection boxes are treated as actual objects, while the potentially ambiguous or falsely identified second detection boxes are marked as indeterminate. The system does not completely ignore the second detection box, but rather tracks it separately as an object with lower confidence, thus allowing for further analysis or verification. Advantageously, the risk of overcounting due to false alarms or ambiguous detections can be reduced. This separation allows for a primary focus on high-confidence objects while still monitoring lower-confidence objects without erroneously overstating the object count.

[0034] In some examples, the method further includes: filtering out the second detection box from an initial set of detection boxes that includes the first detection box, the second detection box, and the third detection box, and using the remaining set of detection boxes in a downstream analysis system.

[0035] In a computing context, the term "downstream" refers to a subsequent process, such as object tracking, object analysis, alarm systems, or other decision-making processes, that depends on the output of earlier steps or systems. In this example, prior to downstream analysis, potentially ambiguous or unreliable second detection boxes are filtered out, ensuring that only the most confident and relevant detection boxes, like the first and third detection boxes, are used for further processing. Advantageously, this approach can help improve the accuracy and efficiency of the downstream system by preventing unreliable or ambiguous detections (such as second detection boxes) from affecting later analysis stages. For example, in object tracking or alarm systems, filtering out potential false alarms reduces the chance of erroneous results such as false alarms or inaccurate object tracking.

[0036] According to a second aspect of this disclosure, the above objective is achieved by a non-transitory computer-readable storage medium having stored thereon instructions for implementing the method according to the first aspect when executed on a processing-capable device.

[0037] According to a third aspect of this disclosure, the above objective is achieved by an anchor-based object detection system configured to identify potential false positive detection boxes in a set of three detection boxes, wherein each of the three detection boxes is detected in the same image frame, wherein each of the three detection boxes is associated with a corresponding predicted intersection-over-union (IoU) score, wherein the predicted IoU score indicates the confidence that the detection box represents an object by the anchor-based object detection system, the anchor-based object detection system being configured to: determine that a second detection box at least partially overlaps with both the first and third detection boxes, wherein the second detection box is located between the first and third detection boxes in the image frame; identify the second detection box as a potential false positive detection box by the following steps: determining that the IoU score of the second detection box is lower than a first threshold score; and determining a first reference point in the first detection box, a corresponding second reference point in the second detection box, and a corresponding third reference point in the third detection box, and determining that the first, second, and third reference points are substantially aligned in the image frame such that the second reference point is within a threshold distance from the alignment defined by the first and third reference points.

[0038] The second and third aspects can generally have the same features and advantages as the first aspect. It should be further noted that, unless otherwise expressly stated, this disclosure relates to all possible combinations of features. Attached Figure Description

[0039] The above and additional objects, features and advantages of this disclosure will be better understood through the following illustrative and non-limiting detailed description of embodiments of the present disclosure with reference to the accompanying drawings, in which the same reference numerals will be used for similar elements, in which:

[0040] Figure 1 The training of anchor frames in an anchor-based object detection system using multiple image frames is illustrated according to an embodiment.

[0041] Figure 2 A first alignment metric is shown between three detection boxes used to identify potential false alarm detection boxes according to an embodiment;

[0042] Figure 3 A second alignment metric between three detection boxes for identifying potential false alarm detection boxes is shown according to an embodiment;

[0043] Figure 4 A flowchart illustrating a method for identifying potential false alarm detection boxes according to an embodiment is shown;

[0044] Figure 5 A system having an object detection system and a downstream analysis system according to an embodiment is shown. Detailed Implementation

[0045] In object detection systems such as Single Shot Detector (SSD) and YOLO (You Only See Once), anchor boxes are widely used on images to detect objects at different scales and aspect ratios. When anchor boxes are densely distributed, it allows for high levels of detection accuracy by covering a wide range of possible object sizes and locations. However, maintaining a dense distribution of anchor boxes requires considerable computational power because the object detection system needs to process a large number of potential detection boxes in each image. To reduce the hardware resources required for this computation and / or to address computational hardware limitations (such as the need for lower-cost or lower-power hardware), the number of anchor boxes can be reduced.

[0046] However, this reduction in anchor boxes introduces challenges during the training process of object detection systems. When the number of anchor boxes is reduced, the object detection system may struggle to distinguish between closely spaced objects. During training, the reduced number of anchor boxes increases the likelihood that a single anchor box will overlap with multiple objects, leading to ambiguous assignments. This ambiguity can cause the object detection system to alternately assign the same anchor box to different objects across different training steps, resulting in intermediate detections. These intermediate detections, occurring due to the sharing of anchor boxes between neighboring objects, reduce detection accuracy and complicate the object detection system's ability to discern precise object boundaries. Therefore, while reducing the number of anchor boxes improves computational efficiency, it also introduces difficulties during training that can be addressed to ensure reliable object detection.

[0047] Figure 1 The example illustrates the reason for intermediate detections during the training of an object detection system. Figure 1 In this example, four images 102, 122, 142, and 162 are used to train the object detection system. Each of images 102, 122, 142, and 162 includes two objects, and as will now be explained, the arrangement of these objects relative to anchor box 108 causes ambiguity during the training process, ultimately leading to intermediate detections.

[0048] The top-left image 102 contains two objects 104 and 106, as well as a single anchor frame 108. It is important to note that in practice, image 108 will include more anchor frames. However, for the sake of simplicity and clarity, only one anchor frame 108 is depicted.

[0049] like Figure 1 As shown, anchor box 108 overlaps with both objects 104 and 106 to a similar degree. This results in the intersection-over-union (IoU) scores of anchor box 108 being similar for both objects. The IoU score is represented by the overlap between anchor box 108 and the dashed rectangle indicating the bounding boxes of the corresponding objects 104 and 106. The similarity of the IoU scores creates ambiguity during training because the system cannot easily determine which object the anchor box should represent. As indicated by arrow 110, in the case of the upper left image 102, the object detection system assigns or encodes anchor box 108 to the left object 104. In this context, "assign" or "encode" means that the object detection system selects anchor box 108 to represent a specific object, in this case the left object 104, thereby adjusting the parameters of the anchor box (bounding box coordinates, class prediction, etc.) to best fit the left object 104.

[0050] Figure 1 The upper right image 122 contains two objects 124 and 126 and an anchor frame 108. In this case, as indicated by arrow 130, the object detection system assigns or encodes the anchor frame 108 to the right object 126.

[0051] Figure 1 The lower left image 142 contains two objects 144 and 146, as well as an anchor box 108. In this case, as indicated by arrow 150, the object detection system assigns or encodes the anchor box 108 to the left object 144.

[0052] Figure 1 The lower right image 162 contains two objects 164 and 166 and an anchor frame 108. In this case, as indicated by arrow 170, the object detection system assigns or encodes the anchor frame 108 to the right object 166.

[0053] like Figure 1As illustrated in the diagram and described above, inconsistencies in anchor box assignment to different objects arise due to similar overlap during training. Specifically, the same anchor box is alternately associated with different objects across multiple training images, potentially reducing the object detection system's ability to accurately determine object boundaries. As understood from the above, these objects may not necessarily be identical across all images, but they may share similar characteristics such as size and position, making it difficult for the system to consistently assign anchor boxes to the correct objects. This problem can primarily occur when two objects with alternating anchor boxes belong to the same category, as this typically means the anchor box fits both objects equally well, making it harder for the system to distinguish between them.

[0054] This issue can cause intermediate detections during the inference phase of an object detection system, where ambiguous or false detection boxes appear between actual objects. Below... Figure 2 and Figure 3 This provides an example of how to identify such intermediate detections by using alignment metrics that analyze the spatial relationships between three partially overlapping detection boxes.

[0055] Figure 2 Image frame 200 is shown, comprising two objects 218 and 220 and three detection boxes 202, 204, and 206. Detection box 202 will be referred to as the first detection box 202. Detection box 206 will be referred to as the second detection box 206. Detection box 204 will be referred to as the third detection box 204.

[0056] To identify potential false alarm detection boxes, it is first determined that the second detection box 206 at least partially overlaps with both the first detection box 202 and the third detection box 204, wherein the second detection box 206 is located between the first detection box 202 and the third detection box 204 in the image frame 200.

[0057] When this second detection box 206 is found, the alignment between the three detection boxes 202, 204, and 206 can be checked.

[0058] It has been discovered by the inventors that if the first detection box 202, the second detection box 204, and the third detection box 206 are substantially aligned, this can indicate that the intermediate detection box 206 is a potential false alarm detection box. There are many ways to determine the alignment metric between the set of three detection boxes 202, 204, and 206. Such a metric can be consistently determined using the corresponding reference points for the respective detection boxes.

[0059] exist Figure 2 In the above, the first reference point 208 is the midpoint of the top edge of the first detection frame 202. Correspondingly, the second reference point 210 is the midpoint of the top edge of the second detection frame 206, and the third reference point 212 is the midpoint of the top edge of the third detection frame 204.

[0060] According to one example, the alignment metric can be determined using a first vector 214 between a first pair of reference points selected from a first reference point, a second reference point, and a third reference point, and a second vector 216 between different second pairs of reference points selected from the first reference point, the second reference point, and the third reference point. This embodiment... Figure 2 The example is shown below.

[0061] exist Figure 2 In the example, the first vector 214 is the vector between the first reference point 208 and the second reference point 210. The second vector 216 is the vector between the second reference point 210 and the third reference point 212.

[0062] In another example, the first vector is the vector between the first reference point 208 and the second reference point 210. The second vector is the vector between the first reference point 208 and the third reference point 212.

[0063] In yet another example, the first vector is the vector between the first reference point 208 and the third reference point 212. The second vector is the vector between the second reference point 210 and the third reference point 212.

[0064] To determine the alignment vector, the cosine of the angle θ between the first and second vectors can be determined: Equation 1 in, = the first vector, and =Second vector.

[0065] Alignment measures can be determined, for example, by the following formula: 1-| Equation 2

[0066] For example, it can be determined that if the absolute value of the cosine of the angle θ between the first and second vectors is less than a threshold value away from 1, then the first, second, and third reference points are essentially aligned in the image frame: 1-| |<ε Equation 3 Where ε = threshold.

[0067] Figure 3 This is another example illustrating how to determine the alignment metric between the set of three detection boxes 202, 204, and 206. Figure 3 The use of Figure 2 The same image frame 200 and the same reference points 208, 210, and 212. According to Figure 3In the example shown, the alignment metric can be determined by determining the distance 304 from the second reference point 210 to the line 302 formed between the first reference point 208 and the third reference point 212.

[0068] If the first reference point 208 is at (x1, y1), the third reference point is at (x3, y3), and the second reference point is at (x2, y2), then the distance d can be determined using the following equation: Equation 4

[0069] The distance d represents the shortest distance from the second reference point to the line bounded by the first and third reference points. Alternatively, if we consider the vector V from the first to the third reference point and take point P as the second reference box, d can also be viewed as the length of the orthogonal projection of P onto V. This projection allows us to determine the alignment of the reference points.

[0070] For example, if d is less than the threshold σ, then the first reference point, the second reference point, and the third reference point are considered to be substantially aligned in the image frame. d<σ Equation 5 Where σ = threshold distance.

[0071] Suitable values ​​for ε and / or σ can be determined, for example, by the following process:

[0072] Run the object detector on the dataset and mark (automatically or manually) detection boxes that overlap with two other detection boxes as potential intermediate detections.

[0073] Then, these labeled detections can be checked and annotated by humans to identify which detections appear to be true intermediate detections.

[0074] Once this operation is completed, angles and / or distances can be analyzed to observe their distribution.

[0075] Based on this analysis, an appropriate threshold (ε and / or σ) can be determined by considering the precision / recall tradeoff and the balance between optimizing accurate detection of intermediate boxes and minimizing false positives or missed detections.

[0076] Figure 4 A flowchart of a method 500 for identifying potential false positive detection boxes in a set of three detection boxes within an anchor-based object detection system is shown by way of example. Method 500 is used to identify potential false positive detection boxes in a set of three detection boxes within an anchor-based object detection system, wherein each of the three detection boxes is detected in the same image frame, and each of the three detection boxes is associated with a corresponding predicted intersection-over-union (IoU) score, wherein the predicted IoU score indicates the confidence of the anchor-based object detection system that the detection box represents an object.

[0077] In object detection systems, the predicted IoU (Intersection over Union) score represents the confidence level of an object detection model's accuracy in determining whether its predicted bounding boxes overlap with objects in an image. It is based on the object detection model's internal understanding, built up through training on the degree to which the detection box is likely to fit the object it should represent. Therefore, the predicted IoU score can be viewed, for example, as an indicator of the potential degree of overlap, measured as the ratio of the intersection area to the combined area of ​​the detection box and the object it represents.

[0078] The predicted IoU score, also known as object degree, reflects the likelihood that the detection box contains an actual object. Throughout training, the object detection model learns these patterns from labeled data, and during inference, the model applies this knowledge to predict the degree to which the detection box will fit the object encountered by the model in new, unseen data.

[0079] In general, the predicted IoU score is a metric reflecting how well the object detection model's bounding box fits the object, which can be used to guide anchor-based object detection systems in deciding whether to retain or discard detections during inference. While effective in many cases, it may sometimes require supplementary analysis to ensure robustness. For example, the IoU score may not be reliable enough when it comes to identifying intermediate detections as described in this paper, and this problem can be addressed, for example, by using... Figure 4 The method shown in the flowchart is used for processing.

[0080] Method 500 includes: determining, in step S502, that a second detection box at least partially overlaps with both a first and a third detection box, wherein the second detection box is located between the first and third detection boxes in an image frame. Method 500 first identifies the overlapping region between the second detection box and both the first and third detection boxes. Then, using detection box coordinates, method 500 verifies whether the second detection box is spatially located between the first and third detection boxes in the frame. This can be done, for example, by checking whether the center of the second detection box falls between the centers of the first and third detection boxes, thereby ensuring that the second box is located between the other two boxes in the image.

[0081] The method then includes identifying, in step S504, that the second detection box is a potential false positive. Depending on the implementation, the identification step S504 may include a variety of numbers of sub-steps. The identification step S504 includes determining, in step S506, that the predicted IoU score of the second detection box is below a first threshold score. Determining an appropriate first threshold score may involve testing and fine-tuning based on empirical data. As described above, one way to identify this threshold is to run an object detector on a dataset of images and apply a manual inspection process. During this inspection, the system annotates the detection boxes with the predicted IoU scores. By analyzing these labeled detections and observing the IoU score distribution of intermediate and true positive detections, the inspector can set a threshold that balances detection accuracy and false positive reduction, typically ensuring reliable detection by minimizing erroneous intermediate detections while incorporating a precision / recall tradeoff.

[0082] The identification step S504 may further include: determining a first reference point in the first detection frame, a corresponding second reference point in the second detection frame, and a corresponding third reference point in the third detection frame, and determining that the first reference point, second reference point, and third reference point are substantially aligned in the image frame, such that the second reference point is within a threshold distance from the alignment defined by the first reference point and the third reference point. This step S508 can be combined as described above. Figure 2 and Figure 3 The above is to achieve this.

[0083] In some examples, identification step S504 may further include: determining that the predicted IoU score of the second detection box in S510 is at least a second threshold score lower than the predicted IoU scores of each detection box in the first and third detection boxes. Similar to the discussion above regarding the first threshold score, the second threshold score can be determined through testing and fine-tuning based on empirical data.

[0084] In some examples, the identification step S504 may further include: determining that S512 is the same as the predicted object category associated with each of the first, second, and third detection boxes.

[0085] In some cases, the detection boxes identified in an image frame can undergo further analysis S514 by a downstream analysis system. This additional analysis S514 can specifically consider second detection boxes that have been identified as potential false positives. Using this information, the downstream analysis system can adjust its analysis S514 accordingly, either by lowering the priority of the second detection box, marking it for more careful examination, or by excluding it from critical decisions to enhance the overall reliability and accuracy of the detection process.

[0086] Figure 5System 400 is illustrated by way of example. System 400 includes an anchor-based object detection system 402 and a downstream analysis system 404. The downstream analysis system 404 uses the output 412 (i.e., detection boxes) from the anchor-based object detection system 402 to perform additional analysis.

[0087] For example, downstream analysis system 404 may include object tracking system 406. Object tracking system 406 may be configured to assign a lower probability of associating an object trajectory with a second detection box compared to the probabilities assigned to the first and third detection boxes. This can be achieved by assigning a higher cost to associating the second detection box with the object trajectory compared to the cost of associating the first or third detection box with the object trajectory. By increasing the cost of associating the second detection box, the tracking system is less likely to link it to object tracking unless a better option is available. This cost-based approach allows object tracking system 406 to prioritize higher-confidence detection boxes (i.e., the first and third detection boxes), thereby reducing the likelihood that potentially false positive detection boxes (the second detection box) will interfere with trajectory association.

[0088] Alternatively or additionally, lower priority can be achieved by the following steps: assigning the second detection box to a lower priority partition of the detection box associated with the object trajectory, and assigning the first and third detection boxes to higher priority partitions of the detection box associated with the object trajectory. The object tracking system 406 can then process these partitions sequentially, first attempting to match the object trajectory with the detection boxes in the higher priority partitions. Only after the high-priority detections have been processed does the object tracking system 406 evaluate the lower priority partitions, where the second detection box resides.

[0089] The object tracking system 406 can be further configured to filter out second detection boxes marked as potential new object trajectories in the object tracking system from the set of detection boxes in the first image frame. In other words, the object tracking system 406 can be configured to exclude second detection boxes from the initial set of candidate detections used to cause new object trajectories, thereby ensuring that only reliable and high-confidence detections (such as the first and third detection boxes) are considered for initializing new trajectories.

[0090] The downstream analysis system 404 may include an object counting system 408. The object counting system 408 may be configured to count the first and third detection boxes as confirmed objects and the second detection box as an unconfirmed object.

[0091] Downstream analysis system 404 may include other analysis systems 410, such as behavior recognition modules or anomaly detection systems, which take the output of object detection system 402 as input for further analysis.

[0092] In some cases, the second detection box is marked as a potential false positive in output 412, allowing the downstream analysis system 404 to take this into account when performing further analysis on the detection boxes from the object detection system 402. This approach enables the downstream system 404 to treat potentially ambiguous detections differently, thereby reducing the impact of false positives in subsequent analysis.

[0093] In other embodiments, object detection system 402 may alternatively filter out second detection boxes from an initial set of detection boxes including first, second, and third detection boxes, and use the remaining set of detection boxes in downstream analysis system 404. Through pre-filtering, object detection system 402 can ensure that only high-confidence detections are used in further analysis, thereby contributing to improved accuracy and reliability of downstream analysis stage 404. The selection of which second detection boxes to label or filter can be implementation-specific and can depend on factors such as system requirements, computing resources, and the intended use of the detection results.

[0094] like Figure 5 The illustrations, including the object detection system 402 and downstream analysis system 404 and their various subsystems, show a functional division within system 400 for processing potential false alarm detection boxes, provided merely for clarity. Components depicted within specific analysis subsystems (e.g., tracking, counting, and anomaly detection) such as object detection system 402 and downstream analysis system 404 are shown as distinct entities to clearly convey the roles and processes involved in processing detection boxes and subsequent analysis. However, it should be understood that the techniques discussed herein can be implemented in various configurations, and the organization of these components can vary based on system architecture and design choices. For example, some of the functions described herein can be integrated into a single module, distributed across multiple subsystems, or implemented through alternative methods that satisfy the same objective. Therefore, Figure 5 The structures described herein are not intended to be restrictive, and any configurations that perform object detection, analysis, and further processing as outlined herein fall within the scope of this disclosure.

[0095] In the example, the methods and techniques described herein, such as method 500, may be implemented using a non-transitory computer-readable storage medium having instructions stored thereon for performing these methods when executed on one or more devices with processing capabilities. By way of example, suitable processors for the program executing the instructions include both general-purpose microprocessors and special-purpose microprocessors, as well as a single processor or one of multiple processors or cores of any kind of computer. The processor can be supplemented by or incorporated into an ASIC (Application-Specific Integrated Circuit).

[0096] The above embodiments should be understood as illustrative examples of the invention. Further embodiments of the invention are contemplated. For example, other methods can be implemented to determine whether the first reference point, the second reference point, and the third reference point are substantially aligned in an image frame. For example, a relatively linear alignment based on the distance between the reference points can be used. When the three points are aligned, the distance between the first and third points should be approximately equal to the sum of the distances from the first point to the second point and from the second point to the third point. A ratio close to 1 indicates alignment, while a deviation from 1 exceeding a threshold indicates misalignment. It should be understood that any feature described with respect to any embodiment can be used alone or in combination with other described features, and can also be used in combination with one or more features of any other embodiment or any combination of any other embodiment. Furthermore, equivalents and modifications not described above may be employed without departing from the scope of the invention as defined by the appended claims.

Claims

1. A method (500) for identifying potential false alarm detection boxes in a set of three detection boxes (202, 204, 206) within an anchor-based object detection system (402), wherein, Detecting each of the three detection boxes in the same image frame (200), wherein each of the three detection boxes is associated with a corresponding predicted intersection-union (IoU) score, wherein the predicted IoU score indicates the confidence of the anchor-based object detection system in representing the object by which the detection box is located, the method comprising: (S502) It is determined that the second detection box (206) at least partially overlaps with both the first detection box (202) and the third detection box (204), wherein the second detection box is located between the first detection box and the third detection box in the image frame; The second detection box is identified as a potential false alarm detection box through the following steps (S504): Determine (S506) that the predicted IoU score of the second detection box is lower than the first threshold score; and (S508) Determine (S508) a first reference point (208) in the first detection frame, a corresponding second reference point (210) in the second detection frame, and a corresponding third reference point (212) in the third detection frame, and determine that the first reference point, the second reference point, and the third reference point are substantially aligned in the image frame such that the second reference point is within a threshold distance from the alignment defined by the first reference point and the third reference point.

2. The method according to claim 1, wherein, Determining that the first reference point, the second reference point, and the third reference point are substantially aligned includes: A first vector (214) is determined between a first pair of reference points selected from the first reference point, the second reference point, and the third reference point, and a second vector (216) is determined between a different second pair of reference points selected from the first reference point, the second reference point, and the third reference point, wherein the first reference point, the second reference point, and the third reference point are substantially aligned in the image frame if the absolute value of the cosine of the angle (θ) between the first vector and the second vector is less than a threshold value of 1.

3. The method according to claim 1, wherein, Determining that the first reference point, the second reference point, and the third reference point are substantially aligned includes: The second reference point is determined to be within a threshold distance from the line formed between the first reference point and the third reference point.

4. The method according to claim 1, wherein, Identifying the second detection box as a potential false alarm is further performed through the following steps: (S510) Determine that the predicted IoU score of the second detection box is at least a second threshold score lower than the predicted IoU score of each detection box in the first and third detection boxes.

5. The method according to claim 1, wherein, Each of the three detection boxes is associated with a predicted object category, wherein identifying the second detection box as a potential false alarm detection is further performed through the following steps: Determine (S512) that the predicted object category associated with each of the first detection box, the second detection box, and the third detection box is the same.

6. The method according to claim 1, wherein, The first reference point, the second reference point, and the third reference point are the midpoints of the top edges of the first detection frame, the second detection frame, and the third detection frame, respectively.

7. The method according to claim 1, wherein, The first reference point, the second reference point, and the third reference point are the center points of the first detection frame, the second detection frame, and the third detection frame, respectively.

8. The method of claim 1, further comprising: The second detection box is assigned a lower probability of being associated with an object trajectory in the object tracking system (406) compared to the probabilities assigned to the first detection box and the third detection box.

9. The method according to claim 8, wherein, Assigning a lower probability includes assigning a higher cost to the second detection box associated with the object trajectory compared to the cost associated with the first or third detection box associated with the object trajectory.

10. The method according to any one of claims 8 and 9, wherein, Assigning a lower probability includes assigning the second detection box to a lower priority partition of the detection box associated with the object trajectory, and assigning the first and third detection boxes to a higher priority partition of the detection box associated with the object trajectory, wherein the partitions are processed sequentially to be associated with the trajectory in the object tracking system.

11. The method of claim 1, further comprising: Filter out the second detection boxes that are marked as potential new object trajectories in the object tracking system from the set of detection boxes in the first image frame.

12. The method of claim 1, further comprising: In the object counting system (408), the first detection box and the third detection box are counted as confirmed objects, and the second detection box is counted as an undetermined object in the object counting system.

13. The method of claim 1, further comprising: The second detection box is filtered out from the initial set of detection boxes including the first detection box, the second detection box and the third detection box, and the remaining set of detection boxes is used in the downstream analysis system (404).

14. A non-transitory computer-readable storage medium having stored thereon instructions for implementing the method according to any one of claims 1 to 13 when executed on one or more means having processing capabilities.

15. An anchor-based object detection system (402) configured to identify potential false alarm detection boxes in a set of three detection boxes (202, 204, 206), wherein, Each of the three detection boxes is detected in the same image frame (200), wherein each of the three detection boxes is associated with a corresponding predicted intersection-union (IoU) score, wherein the predicted IoU score indicates the confidence of the anchor-based object detection system of the object represented by the detection box, the anchor-based object detection system being configured to: (S502) It is determined that the second detection box (206) at least partially overlaps with both the first detection box (202) and the third detection box (204), wherein the second detection box is located between the first detection box and the third detection box in the image frame; The second detection box is identified as a potential false alarm detection box through the following steps (S504): Determine (S506) that the IoU score of the second detection box is lower than the first threshold score; and (S508) Determine (S508) a first reference point (208) in the first detection frame, a corresponding second reference point (210) in the second detection frame, and a corresponding third reference point (212) in the third detection frame, and determine that the first reference point, the second reference point, and the third reference point are substantially aligned in the image frame such that the second reference point is within a threshold distance from the alignment defined by the first reference point and the third reference point.

Images