TWI442326B - Image classification method and image classification system - Google Patents

Description

Image recognition method and image recognition system

The present invention relates to an image recognition method and an image recognition system, and in particular to an image recognition method and image recognition system having a dynamic adjustment feature extraction function and considering the influence of identification timing.

The design and invention of the airbag has effectively reduced the damage caused by traffic accidents, but the airbag blasting itself also causes certain damage to users (such as driving and passengers). Early airbags did not design different bursts of compression for different objects of protection, often resulting in blast damage or inadequate protection. In view of this, many industry and academic institutions are currently developing sensors to detect the driving state, in order to properly control the airbag burst and avoid blast damage.

Currently the most commonly used sensors are weight or pressure sensors, ultrasonic sensors, and camera sensors. The weight or pressure sensor mainly senses the weight or pressure on the seat to identify the driving state, as the inflation speed and force when the airbag pops open, but the most important function of such sensors is used in sensing. The presence or absence of an object does not calculate the distance between the object and the safety. The ultrasonic sensor mainly measures the distance between the target and the object through ultrasonic transmission and reception, and can be applied to accurately detect the distance between the passenger and the airbag. However, the disadvantage of the weight sensor and the weight sensor cannot further the existing object. Classification.

In order to effectively overcome the above difficulties, in order to achieve effective identification and detection of driving state, many computer image recognition technologies have been proposed. Generally, it can be classified into two types: monocular-vision occupant classification and stereo-vision occupant classification.

Monocular vision recognition technology uses a single camera technology to identify the driver's seat state. It can be divided into a template matching method and a machine learning based classifier according to the difference in technology used. ). The identification technology of the template comparison mainly collects the image database of each occupant class, and constructs a corresponding template model for each image, and then calculates by the algorithm of the pattern comparison. The most similar pattern to the current image is used to achieve image classification and identification.

However, the pattern comparison technique is mainly based on the assumption that the patterns are mutually independent. Therefore, in order to obtain a higher recognition rate, sufficient patterns must be collected for different types, but this will greatly reduce image recognition. efficacy. The classification technology based on machine learning is mainly through the theoretical framework of machine learning to obtain classifiers of various categories for classifying different types.

Stereoscopic vision recognition technology mainly imitates human vision to achieve stereoscopic vision. The main principle is to use two cameras arranged at different distances to sense three-dimensional objects, and simultaneously acquire image of the object and calculate the distance of the object. Therefore, it can provide more accurate identification features than the monocular vision recognition technology.

Regardless of whether monocular or stereoscopic vision is applied to passenger state recognition, it can be roughly divided into three steps: foreground segmentation, feature extraction, and occupant classification. However, due to the influence of light fluctuations in the environmental scene, the brightness and color of the image sensed by the camera are not fixed; on the other hand, the background may also have different views over time, especially in the actual application of the vehicle, the external environment. The scene change is more complicated, so that the prior art on the passenger state recognition can not effectively overcome the influence of the environment scene on the recognition over time.

In view of the above, one aspect of the present invention is to provide an image recognition method, which has the function of dynamically adjusting feature extraction and considering the influence of identification timing, can improve the recognition rate, and solve the problems encountered in the prior art.

The image recognition method comprises the following steps: (a) capturing a current image of a target area; and (b) calculating a first effective identification area corresponding to one of the first classification criteria, and calculating a corresponding one of the current images. a feature vector; (c) calculating, according to one of the first classification criteria, the first weight distribution, the first feature vector, a state transition rule, and a previous state of the target region, the target region for the first classification criterion a first state probability; and (d) determining a current state of one of the target regions based on the first state probability.

In actual application, there will be more than one classifier referenced, so the foregoing step (b) further includes a second effective identification area according to one of the corresponding second classification criteria to calculate one of the corresponding ones in the current image. a second feature vector, the step (c) further comprising: calculating a second classification according to a second weight distribution corresponding to the second classification criterion, the second feature vector, a state transition rule, and a previous state of the target region The second state probability of one of the target regions of the criterion, and the step (d) is to determine the current state of the target region based on the first state probability and the second state probability.

Since the image recognition method of the present invention uses the state transition rule, when determining the current state of the target region, a state before the target region can be considered to avoid irrationality of state transition.

In addition, the image recognition method of the present invention utilizes a Rao-Blackwellised particle filter to dynamically adjust the first effective identification region and the first weight distribution to overcome the influence of the environment scene on the identification over time, and may also allow for the capture. The image has variations in translation, orientation, and scale.

Another aspect of the present invention is to provide an image recognition system for implementing the image recognition method of the present invention. The image recognition system comprises an image capture unit, a storage unit and a data processing unit. The image capturing unit is configured to capture the current image of the target area. The storage unit is configured to store the current image, the first effective identification area corresponding to the first classification criterion, the first weight distribution, the state transition rule, the previous state of the target area, and other temporary storage in the identification process. data. The data processing unit is electrically connected to the image capturing unit and the storage unit for performing various calculations and determination operations required by the image recognition method.

Therefore, the image recognition method and the image recognition system of the present invention can overcome the influence of the environment scene on the recognition over time, and avoid the irrationality of the state transition, thereby improving the recognition rate and overcoming the problems of the prior art.

The advantages and spirit of the present invention will be further understood from the following detailed description of the invention.

Referring to FIG. 1 and FIG. 2, FIG. 1 is a flow chart showing an image recognition method according to an embodiment of the present invention, and FIG. 2 is a functional block diagram of the image recognition system 1 according to the specific embodiment. According to the specific embodiment, the image recognition method is applied to the passenger state recognition. Therefore, the image recognition system 1 is directly coupled to the computer of the vehicle. In other words, the data processing unit 122 and the storage unit 124 of the image recognition system 1 can be regarded as vehicles. One part of the computer 12, and the image capturing unit 14 (for example, a camera) of the image recognition system 1 is installed in the vehicle and connected to the vehicle computer 12.

According to the specific embodiment, the image recognition method first captures a current image of a target area by the image capturing unit 14, as shown in step S102; wherein the target area refers to the area where the seat in the vehicle is located (refer to FIG. 3). Then, the corresponding feature vector in the current image is calculated by the data processing unit 122 according to the effective identification area corresponding to the different classification criteria, as shown in step S104. The images captured in the foregoing and the data required for the calculation can be stored in the storage unit 124, and the subsequent ones are the same.

Since the image characteristics are different for different passenger states, the image area that is important for recognition is not the entire captured image. Therefore, there are different regions of interest corresponding to the identification of different passenger states. That is, there are different identification effective areas based on different classification criteria, and the probability of each passenger state can be judged accordingly. According to this embodiment, the state of the passenger (i.e., the state of the target area) may be an unattended state, an item placement state, a back-to-baby seat state, a forward child seat state, a child ride state, and an adult ride state.

The different identification effective areas are established based on maximizing the difference of images obtained by each passenger state, and the determination of the effective area and the selection of feature points are mainly obtained through a machine learning algorithm. Please also refer to FIG. 3, which is a schematic diagram showing the identification effective area of each passenger state. Before implementing the identification, a training program may be implemented to determine each identified effective area, which first collects a sample image of a considerable number of passenger states, and uses Hausdorff Distance to calculate the collective distance corresponding to each identified effective area, and adjusts each identification effectively. The area, the set distance is repeatedly calculated to determine the information required for each identified effective area in the subsequent identification procedure. According to the specific embodiment, FIG. 3 shows the identified effective areas determined in the training program: the item placement state R1, the back-to-baby seat state R2, the forward child seat state R3, the child ride state R4, and the adult ride state R5. However, the invention is not limited thereto.

In addition, the identification effective area is represented by Tchebichef Moment to form a feature vector, and based on the need of real-time identification, the Adaboost algorithm is used to reduce the dimension of the feature vector, and the weight distribution of the corresponding feature vector is calculated simultaneously. It should be noted that the feature vector calculated in step S104 is also as described above, and is also represented by Tchebichef Moment, and only the element having importance is taken out as a constituent element of the feature vector.

According to the specific embodiment, the image recognition method then uses the data processing unit 122 to calculate the weight classification according to the respective classification criteria, the calculated feature vector, a state transition rule, and a previous state of the target region. The state probability of the target area is as shown in step S106. The state transition rule is to limit the transition of the passenger state by probability, that is, the transition probability between the passenger states to provide identification of the dependence of adjacent time points, which can be a finite state machine. ) to describe.

As shown in Figure 4, based on the actual phenomenon, in addition to the unoccupied state, the transition between the other states (back to the baby seat state, the forward child seat state, the child ride state, and the adult ride state) must pass the unattended state, and Based on time continuity and state dependencies, each state has a greater probability of maintaining the original state, and the probability of direct state transition between states (except for the unoccupied state) is zero. Therefore, considering the identification of this state transition rule can be more consistent with the actual passenger state change to improve the recognition accuracy.

Finally, the image recognition method then uses the state probability calculated by the data processing unit 122 according to the foregoing respective classification criteria. In principle, the higher probability is used to determine the current state of the target region, as shown in step S108.

In addition, since the background environment of the target area is complicated to change and the light and dark difference is severe, and the appearance of the passenger is very diverse, the image capturing unit 14 installed may cause the shooting angle due to vibration or other reasons. Different from the preset value, the valid identification area related parameters set in the training program may no longer be applicable.

Therefore, the image recognition method of the present invention improves the stability and accuracy of the ride state recognition, and the time axis of the passenger changes in the image position, so that all the information is integrated into a shell network architecture, and Rao-Blackwellised particle filter is utilized. Rao-Blackwellised Particle Filtering (RBPF) dynamically adjusts the effective identification area, feature vector and its corresponding weight distribution. Therefore, even if the shooting angle of the image capturing unit 14 changes during use, resulting in a change in translation, orientation, and scale, it is still required to extract the captured image. Effectively identify the image of the area and perform the aforementioned steps.

In addition, based on the foregoing description, integrating the training program and the identification program before identification can form a flow chart of the technical architecture of the present case, as shown in FIG. The classifier is a general term for information including classification criteria, effective identification area and weight distribution; the measurement model provides judgment information based directly on the classifier; the RBPF inference structure is an integrated measurement model and state transition rule. To determine the passenger status.

As described in the foregoing embodiments, the image recognition method and the image recognition system of the present invention can overcome the influence of the environment scene on the recognition over time, and avoid the irrationality of the state transition, thereby improving the recognition rate and providing other operations ( For example, the development of a security bag) correct information to overcome the problems of the prior art.

The features and spirit of the present invention will be more apparent from the detailed description of the preferred embodiments. On the contrary, the intention is to cover various modifications and equivalents within the scope of the invention as claimed.

1. . . Image recognition system

12. . . Car computer

14. . . Image capture unit

122. . . Data processing unit

124. . . Storage unit

R1~R5. . . Passenger status

S102~S108. . . step

FIG. 1 is a flow chart showing an image recognition method according to an embodiment of the present invention.

FIG. 2 is a functional block diagram of an image recognition system according to the specific embodiment.

Figure 3 is a schematic diagram showing the identification effective area of each passenger state.

Figure 4 is a schematic diagram showing the finite state machine for each passenger state.

Figure 5 is a flow chart showing the technical architecture of the case.

S102~S108. . . step

Claims (13)

An image recognition method includes the following steps: (a) capturing a current image of a target area; (b) calculating a corresponding one of the current images according to a first effective identification area corresponding to a first classification criterion a feature vector; (c) calculating the target region for the first classification criterion according to a first weight distribution corresponding to the first classification criterion, the first feature vector, a state transition rule, and a previous state of the target region a first state probability, wherein the state transition rule includes a first state, a second state, a first state transition probability from the first state transition to the second state, and a transition from the second state to the second state a first state transition probability, a third state, a transition from the third state to the first state, a third state transition probability, and a transition from the first state to the third state a four-state conversion probability, and the state transition probability of the third state transitioning to the second state or the second state transitioning to the third state is zero; and (d) determining the mesh according to the first state probability One of the current state of the region. The image recognition method according to claim 1, further comprising the step of: reducing the dimension of the first feature vector by using an Adaboost algorithm. The image recognition method of claim 1, wherein the step (b) further comprises: second effective identification according to one of the corresponding second classification criteria a region for calculating a corresponding one of the second feature vectors in the current image, the step (c) further comprising: a second weight distribution corresponding to the second classification criterion, the second feature vector, a state transition rule, and the target a state before the region to calculate a second state probability of the target region for the second classification criterion, and step (d) is to determine the current region of the target region according to the first state probability and the second state probability status. The image recognition method of claim 1, wherein the target area is a vehicle seat, and the current state is selected from a state of no-ride, an item placement state, a back-to-baby seat state, a forward child seat state, One of the group consisting of a child's ride state and an adult ride state. The image recognition method of claim 1, wherein the first state is an unattended state. The image recognition method according to claim 1, further comprising the steps of: in step (b) and step (c), dynamically adjusting the first effective identification region and the first portion by using a Rao-Blackwellised particle filter A weight distribution. The method for image identification according to claim 1, further comprising the steps of: providing a plurality of sample images about the target region; and calculating, according to the plurality of sample images, the step of (a) The first effective identification area of a classification criterion and the first weight distribution. An image recognition system comprising: An image capturing unit for capturing a current image of a target area; a storage unit for storing the current image, a first effective identification area corresponding to a first classification criterion, and a first weight distribution, a state transition rule and a previous state of the target region, wherein the state transition rule includes a first state, a second state, a first state transition probability from the first state transition to the second state, and the second state Transitioning to one of the first states, a second state transition probability, a third state, transitioning from the third state to the first state, a third state transition probability, and transitioning from the first state to the third state a fourth state transition probability, and the state transition probability of the third state transitioning to the second state or the second state transitioning to the third state is zero; and a data processing unit electrically connecting the image Taking the unit and the storage unit, the data processing unit calculates a corresponding first feature vector in the current image according to the first valid identification area corresponding to one of the first classification criteria, Calculating a first state probability of the target region of the first classification criterion according to a first weight distribution corresponding to the first classification criterion, the first feature vector, a state transition rule, and a previous state of the target region, Finally, according to the first state probability, one of the target regions is determined to be in the current state. The image recognition system of claim 8, wherein the data processing unit reduces the dimension of the first feature vector by using an Adaboost algorithm. Such as the image recognition system described in claim 8 of the patent application, wherein the data The processing unit is further configured to calculate a corresponding second eigenvector in the current image according to a second effective identification region corresponding to a second classification criterion, and then according to the second weight distribution corresponding to the second classification criterion, the first a second feature vector, a state transition rule, and a previous state of the target region to calculate a second state probability of the target region for the second classification criterion, and finally, according to the first state probability and the second state probability, The current state of the target area is determined. The image recognition system of claim 8, wherein the target area is a vehicle seat, and the current state is selected from a state of no-ride, an item placement state, a back-to-baby seat state, a forward child seat state, One of the group consisting of a child's ride state and an adult ride state. The image recognition system of claim 8, wherein the first state is an unattended state. The image recognition system of claim 8, wherein the data processing unit dynamically adjusts the first effective identification area and the first weight distribution by using a Rao-Blackwellised particle filter, and stores the same in the storage unit.