Object detection device and confidence threshold adjustment method for object detection
By dynamically adjusting the confidence threshold, using the results of multiple auxiliary identification models to the degree of correlation between the occurrence of objects, the problem of false alarms or missed rate caused by fixed confidence threshold in the prior art is solved, and more accurate object detection is achieved and adaptability.
Patent Information
- Application Number
- CN202411226953.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-07-17
- Filing Date
- 2024-09-03
- Publication Date
- 2024-12-03
AI Technical Summary
When the confidence threshold is set to a fixed value, the existing artificial intelligence object detection model is susceptible to environmental changes, resulting in high false alarm or missed alarm rates, making it difficult to adapt to different application situations.
Dynamically adjust the confidence threshold through the degree of correlation between the results provided by multiple auxiliary identification models (such as optical detection models, temperature detection models) and the occurrence of objects, reduce the probability of misjudgment or misjudgment, and adapt to a variety of application situations.
It improves the judgment accuracy of object detection models, reduces the rate of misjudgment or misjudgment, and maintains good detection performance in different environments and application situations.
Smart Images

Figure CN119068160A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an object detection technology based on artificial intelligence (AI), and in particular to an object detection device and a confidence threshold adjustment method for object detection. Background Art
[0002] Artificial intelligence (AI) is increasingly being used. For example, information (e.g., images) can be obtained from sensors, and then AI recognition can be used to determine whether a specific object appears. In software technology related to AI, the aforementioned object detection model, object classification model, or prediction model (hereinafter referred to as AI model) will provide a corresponding confidence score for the results it produces. If the confidence score is higher, it means that the AI model is more confident in the result, otherwise it is less confident. Therefore, users of AI models will set a confidence threshold to decide whether to adopt the results of the AI model.
[0003] However, if the confidence threshold is set to a fixed value, it will be difficult to adapt to different application scenarios. Some AI models are easily affected by environmental changes. If the confidence threshold is set to a fixed value and the confidence threshold is poorly set at the beginning, the false positive or false negative rate of the object will be too high, resulting in poor performance of the AI model. Summary of the invention
[0004] The present invention provides an object detection device and a confidence threshold adjustment method for object detection, which make the judgment result of the object detection model more accurate, reduce the probability of misjudgment or missed judgment, and can adapt to various application scenarios.
[0005] The object detection device of the present invention includes an optical camera, multiple sensors and a processor. The optical camera is used to sense a real-time image of a detected area. Multiple sensors are used to sense environmental values in the detected area. The processor receives the real-time image and the environmental values. The processor is used to execute an object recognition model and at least one second recognition model, wherein the object recognition model determines whether an object appears in the detected area based on the real-time image to generate a first recognition result and a corresponding first confidence value, and the at least one second recognition model generates at least one second result based on the environmental values. The processor dynamically adjusts the confidence threshold based on the degree of association between the object and the at least one second result and the value of the at least one second result. The processor determines whether to generate an output result based on the first recognition result based on the adjusted confidence threshold and the first confidence value.
[0006] The confidence threshold adjustment method for object detection of the present invention includes the following steps: sensing a real-time image and environmental values of a detected area; executing an object recognition model and at least one second recognition model, wherein the object recognition model determines whether an object appears in the detected area based on the real-time image to generate a first recognition result and a corresponding first confidence value, and the at least one second recognition model generates at least one second result based on the environmental values; dynamically adjusting the confidence threshold based on the degree of correlation between the object and the at least one second result and the value of the at least one second result; and calculating the adjusted confidence threshold based on the initial confidence threshold, the set coefficient upper limit value, one of the positive correlation and the negative correlation, and the second confidence value corresponding to the at least one second result.
[0007] Based on the above, the object detection device and the confidence threshold adjustment method for object detection described in the embodiments of the present invention use the results of the auxiliary recognition model (such as the second recognition model) to dynamically adjust the confidence threshold of the object detection model. This can reduce the probability of misjudgment or missed judgment and can adapt to a variety of application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0008] Figure 1 is a block diagram of an object detection device according to an embodiment of the present invention.
[0009] Figure 2 is a schematic diagram of an object recognition model and a second recognition model according to an embodiment of the present invention.
[0010] Figure 3 is a flow chart of a confidence threshold adjustment method for object detection according to an embodiment of the present invention.
[0011]
Explanation of symbols
[0012] 100: Object detection device
[0013] 107: Hardware Components
[0014] 109: Software Components
[0015] 110: Optical Camera
[0016] 112-1~112-M: Sensor
[0017] 120: Processor
[0018] 130: System memory
[0019] 140: Object Recognition Model
[0020] 150-1~150-N: Second identification model
[0021] 190: Output results
[0022] 210: Real-time image
[0023] 212-1, 212-M: Environmental Values
[0024] 220-1~220-N: Second Result
[0025] S230~S250, S310~S340: Steps of confidence threshold adjustment method
[0026] DT: Confidence Threshold DETAILED DESCRIPTION
[0027] The object detection model based on artificial intelligence needs to set a corresponding confidence threshold. The setting of the confidence threshold requires corresponding professional knowledge to be accurate. If the confidence threshold cannot be set accurately, the false positive or false negative rate of the object detection model may be too high. Another way is to use the trial and error method to continuously adjust the confidence threshold to find a suitable value through experiments under the condition of a large number of false positives or false negatives.
[0028] The embodiment of the present invention utilizes the results of collecting multiple auxiliary recognition models (such as smoke detection model, optical detection model, temperature detection model, etc.), and dynamically adjusts the confidence threshold of the object detection model by using the correlation between these results and the appearance of the object (such as whether these results are positively correlated or negatively correlated with the appearance of the object), so as to make the judgment result of the object detection model more accurate, so that the embodiment of the present invention can adapt to a variety of application scenarios and reduce the probability of misjudgment or missed judgment. The auxiliary recognition model can be judged by artificial intelligence, or it can be obtained by a preset software algorithm to obtain the corresponding result.
[0029] Figure 1 is a block diagram of an object detection device 100 according to an embodiment of the present invention. Figure 1 The object detection device 100 includes an optical camera 110 , a plurality of sensors 112 - 1 to 112 -M (M is a positive integer and M is greater than or equal to 1), a hardware component 107 and a software component 109 .
[0030] The hardware components 107 include a processor 120 and a system memory 130. The processor 120 may be a central processing unit (CPU), a graphics processing unit (GPU), or corresponding hardware. The system memory 130 may be a cache, a hard disk, a random access memory, or corresponding hardware for accessing data.
[0031] The software component 109 may be implemented by a program code stored in the system memory 130 or a corresponding storage medium, and by the processor 120 and the system memory 130 in the hardware component 107. The software component 109 of this embodiment may include system software (e.g., an operating system), an object recognition model 140, and at least one second recognition model 150-1 to 150-N (N is a positive integer and N is greater than or equal to 1). In other words, the processor 120 is used to execute the object recognition model 140 and at least one second recognition model 1501 to 150-N.
[0032] The optical camera 110 is used to sense the real-time image of the detected area. The object recognition model 140 can determine whether a specific object appears in the detected area based on the real-time image to generate a first recognition result, and generate a corresponding first confidence value for the first recognition result. The object of this embodiment can be a flame, a vehicle, etc., which is a public security hazard. The object recognition model 140 of this embodiment can determine whether the aforementioned object appears in the detected area through artificial intelligence or a corresponding algorithm. The first confidence value is used to indicate the confidence level of the object recognition model 140 in the first recognition result.
[0033] Sensors 112-1~112-M are used to sense the environmental values in the detected area. The environmental values may include one or a combination of temperature values, optical transmittance, and suspended particle values. The type of sensor is adjusted accordingly depending on the requirements of the second recognition model for the environmental values. The second recognition models 150-1~150-N may be one or a combination of optical detection models, temperature detection models, and smoke detection models. Each second recognition model 150-1~150-N may generate a corresponding second result according to the corresponding environmental value. The optical detection model may obtain the optical characteristic sensing result through the corresponding sensing value of the far infrared detector; the temperature detection model may obtain the current temperature in the detected area through the corresponding temperature value of the temperature detector; and the smoke detection model may obtain the suspended particle concentration in the detected area through the suspended particle value detected by the sensor. The second recognition models 150-1~150-N of this embodiment may also be referred to as auxiliary recognition models. Those who apply this embodiment may select different recognition models as auxiliary second recognition models according to their needs.
[0034] The processor 120 dynamically adjusts the confidence threshold DT based on the correlation between the aforementioned object and the second result and the value of the aforementioned second result. In addition, the processor 120 determines whether to generate the output result 190 based on the first recognition result based on the confidence threshold DT and the first confidence value generated by the object recognition model 140. For detailed operations, please see Figure 2 and corresponding detailed descriptions.
[0035] Figure 21 is a schematic diagram of an object recognition model 140 and second recognition models 150 - 1 to 150 -N according to an embodiment of the present invention. Figure 1 and Figure 2 , Figure 1 The processor 120 is used to execute the object recognition model 140 and the second recognition models 150-1 to 150-N. The object recognition model 140 generates a first recognition result and a corresponding first confidence value based on the real-time image. The second recognition models 150-1 to 150-N generate at least one second result 220-1 to 220-N according to the environment values 212-1 to 212-M.
[0036] Figure 2 In step S230, Figure 1 The processor 120 dynamically adjusts the confidence threshold based on the degree of association between the object and the second result and the value of the second result. Figure 1 The processor 120 first sets an initial confidence threshold of the object recognition model 140 , and then dynamically adjusts the confidence threshold according to different scenarios. Figure 1 The processor 120 sets corresponding coefficient upper limits for the second identification models 150-1 to 150-N. Figure 1 The processor 120 sets the second recognition models 150-1 to 150-N to be either positively correlated or negatively correlated based on the correlation between the aforementioned object and the second results 220-1 to 220-N. And, Figure 1 The processor 120 calculates an adjusted confidence threshold based on the aforementioned initial confidence threshold, the set coefficient upper limit value, one of the positive correlation and the negative correlation, and the second confidence value corresponding to the second results 220 - 1 ˜ 220 -N.
[0037] Figure 2 In step S240, Figure 1 The processor 120 determines whether an output result can be generated based on the first recognition result based on the adjusted confidence threshold and the first confidence value corresponding to the first recognition result (step S250). In this embodiment, assuming that the adjusted confidence threshold is "0.8", and assuming that the first confidence value is "0.56", since the first confidence value "0.56" is less than the adjusted confidence threshold "0.8", it means that the first recognition result cannot pass the confidence test and no output result will be generated. In contrast, assuming that the first confidence value is "0.91", since the first confidence value "0.91" is greater than the adjusted confidence threshold "0.8", it means that the first recognition result has passed the confidence test, and therefore an output result will be generated based on the first recognition result.
[0038] Here is an example to illustrate Figure 2 Step S230. Please refer to Figure 2In the first embodiment, the detected area is a liquid nitrogen storage plant of a large chemical plant. The object recognition model 140 is an artificial intelligence model for detecting flames, that is, "flame" is the object to be detected. The second recognition models 150-1~150-N are respectively an optical detection model and a temperature detection model, and N is equal to 2. In other words, the second recognition models 150-1~150-N are represented as multiple auxiliary recognition models. Assume that the processor sets the initial confidence threshold of the object recognition model 140 to "0.8".
[0039] Considering the optical and chemical properties of liquid nitrogen stored in the detected area when it evaporates, the upper limit value of the coefficient of the optical detection model is set to "0.2", and the upper limit value of the coefficient of the temperature detection model is set to "0.1"; the correlation degree of the optical detection model is "positive correlation" (that is, when the detection of light changes its value, the probability of the flame appearing is higher), and the correlation degree of the temperature detection model is "positive correlation" (that is, when the ambient temperature is higher, the probability of the flame appearing is higher). The explanation of the upper limit value of the coefficient of the optical detection model "0.2" is as follows: liquid nitrogen occasionally needs to be depressurized during the storage process, and the release of some liquid nitrogen will cause smoke, but these smoke is not smoke from flames, so this embodiment lowers the impact of the optical detection model a little (that is, the upper limit value of the coefficient is "0.2"). In addition, the upper limit value of the coefficient of the temperature detection model is "0.1" as follows. In this embodiment, the influence of the temperature detection model is lowered (that is, the upper limit value of the coefficient is "0.1") because liquid nitrogen will cause a relatively large temperature difference, and liquid nitrogen will not support combustion and is not flammable after being gasified into nitrogen, so the influence on flame detection can be reduced to a very low level.
[0040] Therefore, when the optical detection model has a detection action, the processor normalizes the second result generated by the optical detection model, so as to selectively set the second result to a unit positive value or a unit negative value based on one of the set positive correlation and negative correlation. For example, the normalized value is a unit negative value (i.e., "-1") when the optical detection model has a detection action; and a unit positive value (i.e., "+1") when the optical detection model has no detection action. In addition, the processor normalizes the second result generated by the temperature detection model, and the normalized value is a unit negative value (i.e., "-1") when the temperature detection model has a detection action; and a unit positive value (i.e., "+1") when the temperature detection model has no detection action.
[0041] Therefore, when the optical detection model has a detection action (the normalized value is "-1") and the corresponding second confidence value is 100% (the upper limit value of the coefficient is "0.2"), and when the temperature detection model has no detection action (the normalized value is "+1") and the second confidence value is 100% (the upper limit value of the coefficient is "0.1"), the processor calculates the adjusted confidence threshold based on the initial confidence threshold ("0.8"), the set coefficient upper limits ("0.2" and "0.1" respectively), one of the positive correlation and the negative correlation, and the second confidence value corresponding to the second result ("100%" and "100%" respectively). That is, the initial confidence threshold ("0.8") plus the normalized value ("-1") of the optical detection model multiplied by the second confidence value ("100%") and the coefficient upper limit value "0.2", plus the normalized value ("+1") of the temperature detection model multiplied by the second confidence value ("100%") and the coefficient upper limit value "0.1" as the adjusted confidence threshold (0.8-0.2+0.1="0.7"). In this embodiment, the adjusted confidence threshold is different from the initial confidence threshold.
[0042] In this embodiment, the initial confidence threshold is "0.8" and the adjusted confidence threshold is "0.7". Assuming that the first confidence value is "0.75", since the first confidence value "0.75" is less than the initial confidence threshold "0.8", it means that the first recognition result has not passed the confidence test, so no output result will be generated. However, when the initial confidence threshold is adjusted from "0.8" to "0.7", the first confidence value "0.75" is greater than the adjusted confidence threshold "0.7", indicating that the first recognition result has passed the confidence test. Therefore, the output result can be generated based on the first recognition result.
[0043] In reference Figure 2 In the second embodiment, the detected area is a butadiene storage plant of a large chemical plant. The object recognition model 140 is an artificial intelligence model for detecting flames. The second recognition models 150-1 to 150-N are still optical detection models and temperature detection models. Assume that the processor sets the initial confidence threshold of the object recognition model 140 to "0.8".
[0044] Considering the optical and chemical properties of volatilized butadiene stored in the detected area, the upper limit value of the coefficient of the optical detection model is set to "0.3", and the upper limit value of the coefficient of the temperature detection model is set to "0.3"; the correlation degree of the optical detection model is "positive correlation", and the correlation degree of the temperature detection model is "positive correlation".
[0045] Therefore, when the optical detection model has a detection action (the normalized value is "-1") and the corresponding second confidence value is 100% (the upper limit value of the coefficient is "0.3"), and when the temperature detection model has no detection action (the normalized value is "+1") and the second confidence value is 100% (the upper limit value of the coefficient is "0.3"), the processor calculates the adjusted confidence threshold based on the initial confidence threshold ("0.8"), the set coefficient upper limits ("0.3" and "0.3" respectively), one of the positive correlation and the negative correlation, and the second confidence value corresponding to the second result ("100%" and "100%" respectively). That is, the initial confidence threshold ("0.8") plus the normalized value ("-1") of the optical detection model multiplied by the second confidence value ("100%") and the coefficient upper limit value "0.3", plus the normalized value ("+1") of the temperature detection model multiplied by the second confidence value ("100%") and the coefficient upper limit value "0.3" is used as the adjusted confidence threshold (0.8-0.3+0.3="0.8"). In this way, although the optical detection model has a detection action, since the temperature detection model has no detection action, it is better to maintain the original initial confidence threshold) "0.8") for the probability of the object recognition model 140 detecting the flame. In this embodiment, the adjusted confidence threshold is the same as the initial confidence threshold.
[0046] In reference Figure 2 In the third embodiment, the detected area is a vinyl chloride storage plant of a large chemical plant. The object recognition model 140 is an artificial intelligence model for detecting flames. The second recognition models 150-1 to 150-N are still optical detection models and temperature detection models. Assume that the processor sets the initial confidence threshold of the object recognition model 140 to "0.8".
[0047] Considering the optical and chemical properties of vinyl chloride stored in the detected area when it volatilizes, the upper limit of the coefficient of the optical detection model is set to "0.3", and the upper limit of the coefficient of the temperature detection model is set to "0.1"; the correlation degree of the optical detection model is "positive correlation", and the correlation degree of the temperature detection model is "positive correlation". The reason why the upper limit of the coefficient of the temperature detection model is set to "0.1" here is that the heat of vaporization of vinyl chloride is low and will not cause any temperature change. Therefore, the second result generated by the temperature detection model is not very useful for reference.
[0048] Therefore, when the optical detection model has a detection action (the normalized value is "-1") and the corresponding second confidence value is 100% (the upper limit value of the coefficient is "0.3"), and when the temperature detection model has no detection action (the normalized value is "+1") and the second confidence value is 100% (the upper limit value of the coefficient is "0.1"), the processor calculates the adjusted confidence threshold based on the initial confidence threshold ("0.8"), the set coefficient upper limits ("0.3" and "0.1" respectively), one of the positive correlation and the negative correlation, and the second confidence value corresponding to the second result ("100%" and "100%" respectively). That is, the initial confidence threshold ("0.8") plus the normalized value ("-1") of the optical detection model multiplied by the second confidence value ("100%") and the coefficient upper limit value "0.3", plus the normalized value ("+1") of the temperature detection model multiplied by the second confidence value ("100%") and the coefficient upper limit value "0.1" as the adjusted confidence threshold (0.8-0.3+0.1="0.6"). In this embodiment, the adjusted confidence threshold is different from the initial confidence threshold.
[0049] In this embodiment, the initial confidence threshold is "0.8" and the adjusted confidence threshold is "0.6". Assuming that the first confidence value is "0.7", since the first confidence value "0.7" is less than the initial confidence threshold "0.8", it means that the first recognition result has not passed the confidence test, so no output result will be generated. However, when the initial confidence threshold is adjusted from "0.8" to "0.6", the first confidence value "0.7" is greater than the adjusted confidence threshold "0.6", indicating that the first recognition result has passed the confidence test. Therefore, the output result can be generated based on the first recognition result.
[0050] In reference Figure 2 In the fourth embodiment, the detected area is the chimney area of the combustion tower. The object recognition model 140 is an artificial intelligence model for detecting air pollution smoke. In other words, the object recognition model 140 is a smoke detection model based on RGB images. The second recognition models 150-1 to 150-N are RGB image flame detection models and temperature detection models. Assume that the processor sets the initial confidence threshold of the object recognition model 140 to "0.8".
[0051] Considering the causes and characteristics of air pollution smoke in the detected area, the upper limit value of the coefficient of the RGB image flame detection model is set to "0.3", and the upper limit value of the coefficient of the temperature detection model is set to "0.3"; the correlation degree of the optical detection model is "positive correlation", and the correlation degree of the temperature detection model is "positive correlation".
[0052] Therefore, when the RGB image flame model has a detection action (the normalized value is "-1") and the corresponding second confidence value is 100% (the upper limit value of the coefficient is "0.3"), and when the temperature detection model has a detection action (the normalized value is "-1") and the second confidence value is 100% (the upper limit value of the coefficient is "0.3"), the processor calculates the adjusted confidence threshold based on the initial confidence threshold ("0.8"), the set coefficient upper limits ("0.3" and "0.3" respectively), one of the positive correlation and the negative correlation, and the second confidence value corresponding to the second result ("100%" and "100%" respectively). That is, the initial confidence threshold ("0.8") plus the normalized value ("-1") of the optical detection model multiplied by the second confidence value ("100%") and the coefficient upper limit value "0.3", plus the normalized value ("-1") of the temperature detection model multiplied by the second confidence value ("100%") and the coefficient upper limit value "0.3" as the adjusted confidence threshold (0.8-0.3-0.3="0.2"). In this way, since both the RGB image flame model and the temperature detection model have detection actions, it indicates that the object recognition model 140 has a greater probability of detecting air pollution smoke, so the confidence threshold will be dynamically lowered. In this embodiment, the adjusted confidence threshold is different from the initial confidence threshold.
[0053] In this embodiment, the initial confidence threshold is "0.8" and the adjusted confidence threshold is "0.2". Assuming that the first confidence value is "0.3", since the first confidence value "0.3" is less than the initial confidence threshold "0.8", it means that the first recognition result has not passed the confidence test, so no output result will be generated. However, when the initial confidence threshold is adjusted from "0.8" to "0.2", the first confidence value "0.3" is greater than the adjusted confidence threshold "0.2", indicating that the first recognition result has passed the confidence test. Therefore, the output result can be generated based on the first recognition result.
[0054] Figure 3 is a flow chart of a confidence threshold adjustment method for object detection according to an embodiment of the present invention. Figure 3 The confidence threshold adjustment method is applicable to the above Figure 1 and Figure 2Corresponding object detection device 100. In step S310, the processor uses an optical camera and multiple sensors to respectively sense the real-time image and environmental values of the detected area. In step S320, the processor executes an object recognition model and at least one second recognition model. The aforementioned object recognition model determines whether an object appears in the detected area based on the real-time image to generate a first recognition result and a corresponding first confidence value. The aforementioned at least one second recognition model generates at least one second result based on the environmental value. In step S330, the processor dynamically adjusts the confidence threshold based on the degree of association between the aforementioned object and the aforementioned second result and the value of the aforementioned second result. In step S340, the processor determines whether to generate an output result based on the first recognition result generated by the object recognition model based on the adjusted confidence threshold and the aforementioned first confidence value. Figure 3 The detailed process and implementation of each step in the confidence threshold adjustment method can be referred to the above embodiments.
[0055] In summary, the object detection device and the confidence threshold adjustment method for object detection described in the embodiments of the present invention utilize the results of the auxiliary recognition model (e.g., the second recognition model) to dynamically adjust the confidence threshold of the object detection model, which can reduce the probability of misjudgment or missed judgment and can adapt to a variety of application scenarios. In addition, when dynamically adjusting the confidence threshold, the interaction results of multiple auxiliary recognition models (e.g., the second recognition model) can be referred to, which can also avoid the problem of a single auxiliary recognition model excessively affecting the confidence threshold.
[0056] Although the present invention has been disclosed as above by way of embodiments, it is not intended to limit the present invention. Those skilled in the art may make slight changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be determined by the scope defined in the attached claims.
Claims
1. An object detection device, comprising: An optical camera for sensing a real-time image of the inspected area; A plurality of sensors for sensing environmental values in the detected area; as well as A processor receives the real-time image and the environmental value, wherein the processor is used to execute an object recognition model and at least one second recognition model, wherein the object recognition model determines whether an object appears in the detected area based on the real-time image to generate a first recognition result and a corresponding first confidence value, and the at least one second recognition model generates at least one second result according to the environmental value. The processor dynamically adjusts the confidence threshold based on the degree of association between the object and the at least one second result and the value of the at least one second result, and The processor determines whether to generate an output result according to the first recognition result based on the adjusted confidence threshold and the first confidence value.
2. The object detection device according to claim 1, wherein the at least one second recognition model comprises a plurality of auxiliary recognition models, The processor is also used to: The confidence threshold is dynamically adjusted based on a plurality of second results generated by the auxiliary identification model and a coefficient upper limit value corresponding to each auxiliary identification model.
3. The object detection device according to claim 1, wherein the processor is used to: Setting an initial confidence threshold of the object recognition model; Setting corresponding coefficient upper limits for the at least one second identification model; Based on the correlation degree between the object and the at least one second result, the at least one second recognition model is set to be either positively correlated or negatively correlated; as well as The adjusted confidence threshold is calculated based on the initial confidence threshold, the set coefficient upper limit value, one of the positive correlation and the negative correlation, and a second confidence value corresponding to the at least one second result. The object detection device as claimed in claim 3 , wherein the value of the initial confidence threshold is different from the value of the adjusted confidence threshold. 5 . The object detection device as claimed in claim 3 , wherein the value of the initial confidence threshold is the same as the value of the adjusted confidence threshold.
6. The object detection device according to claim 3, wherein the processor is further configured to: The at least one second result is normalized to selectively set the at least one second result to a unit positive value or a unit negative value based on one of the set positive correlation and the set negative correlation. 7 . The object detection device as claimed in claim 1 , wherein the at least one second recognition model is one of an optical detection model, a temperature detection model, a smoke detection model, or a combination thereof. 8 . The object detection device as claimed in claim 1 , wherein the environmental value comprises one of a temperature value, an optical transmittance, and a suspended particle value, or a combination thereof.
9. A confidence threshold adjustment method for object detection, comprising: Sense real-time images and environmental values of the detected area; Executing an object recognition model and at least one second recognition model, wherein the object recognition model determines whether an object appears in the detected area based on the real-time image to generate a first recognition result and a corresponding first confidence value, and the at least one second recognition model generates at least one second result according to the environmental value; dynamically adjusting the confidence threshold based on the degree of association between the object and the at least one second result and the value of the at least one second result; as well as Based on the adjusted confidence threshold and the first confidence value, it is determined whether to generate an output result according to the first recognition result.
10. The confidence threshold adjustment method according to claim 9, wherein the at least one second identification model comprises a plurality of auxiliary identification models, in, Based on the degree of association between the object and the at least one second result and the value of the at least one second result, the step of dynamically adjusting the confidence threshold comprises: The confidence threshold is dynamically adjusted based on a plurality of second results generated by the auxiliary identification model and a coefficient upper limit value corresponding to each auxiliary identification model.
11. The confidence threshold adjustment method of claim 9, wherein the step of dynamically adjusting the confidence threshold based on the degree of association between the object and the at least one second result and the value of the at least one second result comprises: Setting an initial confidence threshold of the object recognition model; Setting corresponding coefficient upper limits for the at least one second identification model; Based on the correlation degree between the object and the at least one second result, the at least one second recognition model is set to be either positively correlated or negatively correlated; as well as The adjusted confidence threshold is calculated based on the initial confidence threshold, the set coefficient upper limit value, one of the positive correlation and the negative correlation, and a second confidence value corresponding to the at least one second result. 12 . The confidence threshold adjustment method according to claim 11 , wherein the value of the initial confidence threshold is different from the value of the adjusted confidence threshold. 13 . The confidence threshold adjustment method according to claim 11 , wherein the value of the initial confidence threshold is the same as the value of the adjusted confidence threshold.
14. The confidence threshold adjustment method according to claim 11, wherein the step of setting the at least one second recognition model to be either positively correlated or negatively correlated based on the degree of correlation between the object and the at least one second result comprises: The at least one second result is normalized to selectively set the at least one second result to a unit positive value or a unit negative value based on one of the set positive correlation and the set negative correlation. 15 . The confidence threshold adjustment method according to claim 9 , wherein the at least one second recognition model is one of an optical detection model, a temperature detection model, a smoke detection model, or a combination thereof. 16 . The confidence threshold adjustment method as claimed in claim 9 , wherein the environmental value comprises one of a temperature value, an optical transmittance, and a suspended particle value, or a combination thereof.
