Intelligent camera

Abstract

The invention relates to the technical field of cameras, in particular to an intelligent camera which comprises a shell and a main board, the main board is installed in the shell, a communication module is further installed in the shell, a data connecting hole is formed in the outer portion of the shell, and a loudspeaker, a lens, an infrared lamp and a light supplementing lamp are arranged on the shell. The main board is electrically connected with the communication module, the loudspeaker, the lens, the infrared lamp and the light supplement lamp, the main board adopts a salary algorithm, the salary algorithm adopts an XHModelv4 model, and the XHModelv4 model comprises an image extraction layer, an aggregation feature layer and a classification positioning layer. The system is composed of a camera and a salary algorithm, the camera provides an ultra-definition lens and high-performance calculation and parallel processing capabilities, and strong support is provided for operation of the salary algorithm; according to the salary algorithm, the computing power provided by a built-in mainboard of the camera is utilized, and a neural network is deeply trained and optimized, so that the strong recognition and judgment capability is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of cameras, in particular to a smart camera. Background Art

[0002] A camera is a device that captures images and converts optical image signals into electrical signals. When we shoot an object, the light reflected from the object is collected by the camera lens and focused on the light-receiving surface of the camera device (such as the target surface of the camera tube). The camera device then converts the light into electrical energy, which is a "video signal". The photoelectric signal is very weak and needs to be amplified by a pre-amplifier circuit, and then processed and adjusted by various circuits. The final standard signal can be sent to a recording medium such as a video recorder for recording, or transmitted through a transmission system or sent to a monitor for display.

[0003] With the development of computer technology, many algorithms are combined with cameras to facilitate people's use. Although existing cameras can identify people or objects, there are many false detections and missed detections, which cannot meet people's usage needs. Summary of the invention

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A smart camera comprises a shell and a mainboard, wherein the mainboard is installed inside the shell and a communication module is also installed inside the shell, a data connection hole is arranged outside the shell, a speaker, a lens, an infrared lamp and a fill light are arranged on the shell, the mainboard is electrically connected to the communication module, the speaker, the lens, the infrared lamp and the fill light respectively, the mainboard adopts a Xinhuo algorithm, the Xinhuo algorithm adopts an XHModelv4 model, the XHModelv4 model comprises an image extraction layer, an aggregation feature layer and a classification and positioning layer, and the XHModelv4 model is optimized as follows: 1. The XHModelv4 model supports an anchor-free method, so that the detection process is more flexible; 2. The XHModelv4 model improves the traditional NMS algorithm, and converts the traditional NMS algorithm into an adaptive NMS algorithm by adaptively adjusting the threshold, so as to reduce false detection and missed detection. The adaptive NMS algorithm in the present invention is called a Soft-NMS algorithm. When calculating IoU, the Soft-NMS algorithm allows the score of overlapping frames to be gradually reduced instead of being directly suppressed, that is, frames with scores lower than the threshold are gradually removed instead of being eliminated across the board, which means that even if overlapping frames exist, the scores will not be completely erased, thereby retaining more information. Therefore, the XHModelv4 model shows higher efficiency and accuracy in the field of target detection, and its innovative structural design and module combination make it have broad potential in practical applications. By using more advanced network components and training strategies, the XHModelv4 model has made significant progress in accuracy and real-time performance, and is suitable for a variety of detection tasks.

[0006] As a further solution of the present invention: the XHModelv4 model adopts CSPNet (Cross Stage PartialNetwork) as the backbone layer. This backbone layer is the image extraction layer, which is a network for extracting image features. The image extraction layer using CSPNet includes four parts. The first part includes two Conv modules and one C2f module. The second and third parts both include a Conv module and a C2f module. The fourth part includes a Conv module, a C2f module and a SPPF module.

[0007] As a further solution of the present invention: the aggregation feature layer is a network for aggregating features. As a Neck layer, the multi-scale feature map from the image extraction layer is merged to enhance the feature representation capability. The aggregation feature layer in the present invention is different from the aggregation feature layer of the traditional convolutional neural network. The aggregation feature layer of the present invention is not an independent part, but is included in the classification and positioning layer. Another special feature is that multiple modules are used to combine the PANet (Path Aggregation Network) from the Backbone layer. It combines features through bottom-up paths and top-down paths to improve target detection accuracy. The aggregation feature layer includes an upsampling layer (Upsample). The upsampling layer performs upsampling operations to increase the spatial resolution of the feature map.

[0008] As a further solution of the present invention: the classification and positioning layer serves as the Head layer, which is used to perform target classification and positioning, and includes a detection head and a classification head. The detection head is responsible for predicting the bounding box regression value of each anchor box and the confidence of the target existence; the classification head is used to classify each feature map by average pooling, and output the probability distribution of each category. The classification and positioning layer includes a prediction layer (Detect), which is responsible for outputting the final detection result, including three outputs: a bounding box (x, y, w, h), confidence, and probabilities of each category; the classification Loss uses BCE (binary cross entropy), and the bounding box regression Loss uses CIOU. BCE and CIOU are both conventional algorithms in this field, and will not be described in detail here.

[0009] As a further solution of the present invention: the Conv module is a standard convolution module, including a Conv2d layer, a BatchNorm2d layer and a SiLU activation function, which are conventional techniques in the art and will not be described in detail here.

[0010] As a further solution of the present invention: the core idea of ​​the C2f module is to divide the input feature map into two parts, process them on different paths, and then merge the results. The C2f module includes an input unit, a feature extraction unit, a feature separation unit, a branch processing unit and a feature merging unit. The input unit is used to receive the feature map from the previous layer. The feature extraction unit extracts the basic features of the input image through an initial standard convolution module. The feature separation unit divides the feature map into two parts (for example, a left half and a right half).

[0011] As a further solution of the present invention: the branch processing unit includes a main branch unit and an auxiliary branch unit, the main branch unit performs standard convolution and processing on a part thereof, and the auxiliary branch unit performs independent processing on the other part, usually including convolution, activation and other operations. At this time, the feature merging unit merges the output feature maps of the main branch unit and the auxiliary branch unit to form the final output, which can improve information flow and feature representation capabilities, which helps to reduce gradient disappearance and improve the training stability of the model.

[0012] As a further solution of the present invention: The SPPF module is designed to enhance the model's ability in multi-scale feature extraction. It helps the network to better handle targets of different sizes and shapes through pooling operations of different sizes. The SPPF module includes an input feature map unit, a preliminary feature extraction unit, a pyramid pooling unit, and a convolution unit. The input feature map unit is used to receive the feature map from the previous layer, the preliminary feature extraction unit performs 1x1Conv convolution processing, and the pyramid pooling unit has the following functions: using maximum pooling (MaxPool2d) for operation; using different pooling sizes (1x1, 5x5, 9x9) to operate on the same input feature map. This multi-scale pooling can capture features at different levels; splicing the output feature maps of each pooling operation. The convolution unit further processes the spliced ​​output feature map using a 1x1 Conv convolution layer and outputs the feature map.

[0013] As a further solution of the present invention: the lens adopts an ultra-clear lens of 8 million pixels, which can clearly and accurately capture the required video.

[0014] As a further solution of the present invention: the data connection hole includes a WIFI hole, a 4G hole and a tail line hole, which are suitable for different working conditions.

[0015] As a further solution of the present invention: a heat sink is also connected to the mainboard to dissipate heat for the mainboard at all times, thereby ensuring that the whole can work stably for a long time.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] The present invention consists of two parts: a camera and a Xinhuo algorithm. The camera provides ultra-clear lenses, high-performance computing and parallel processing capabilities, providing strong support for the operation of the Xinhuo algorithm. The Xinhuo algorithm utilizes the computing power provided by the camera's built-in motherboard, and achieves powerful recognition and judgment capabilities through deep training and optimization of neural networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a front view of the smart camera in an embodiment of the present invention.

[0019] Figure 2It is a rear view of the smart camera in an embodiment of the present invention.

[0020] Figure 3 FIG. 1 is a bottom view of the smart camera according to an embodiment of the present invention.

[0021] Figure 4 FIG. 4 is an internal diagram of a smart camera according to an embodiment of the present invention.

[0022] Figure 5 FIG. 4 is a stereoscopic diagram of a smart camera according to an embodiment of the present invention.

[0023] Figure 6 This is a schematic diagram of the structure of the XHModelv4 model used in the smart camera in an embodiment of the present invention.

[0024] Figure 7 Schematic diagram of the working principle of the C2f module in the smart camera in an embodiment of the present invention.

[0025] Figure 8 Schematic diagram of the working principle of the SPPF module in the smart camera in an embodiment of the present invention.

[0026] In the picture: 1- lens, 2- infrared light, 3- fill light, 4- WIFI port, 5- 4G port, 6- tail line port, 7- speaker, 8- communication module, 9- wire, 10- mainboard, 11- heat sink, 12- casing. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The product of the present invention can realize the following functions: 1. AI real-time monitoring and early warning. The identification process is as follows: set the monitoring area - real-time analysis of the camera monitoring screen - capture the screen after discovering an abnormal event - upload the video and captured pictures to the built-in information platform - alarm prompts and record queries on the information platform. 2. Event query and tracing. After receiving the on-site early warning report event, the early warning event can be displayed in the form of a pop-up window, etc., and the function of retrieving event pictures and videos according to date, camera location, etc. is provided, which plays the role of event early warning and tracing record query and storage.

[0029] The configuration of the product of the present invention is flexible, and the required algorithm, effective time and area, etc. can be set through the built-in information platform of the camera; and the Xinhuo algorithm covers all the algorithms necessary for safety supervision, and can realize the following functions: 1. Human attribute recognition: face recognition, specific work clothes detection, designated color work clothes detection; safety helmet detection, reflective clothing detection, chef hat detection, etc. 2. Human behavior feature recognition: dynamic crowd statistics; leaving the post detection, sleeping on the post detection; smoking detection, phone call detection; climbing over the fence detection, restricted area intrusion detection; gathering detection, detention detection, fall detection, etc. 3. Environmental attribute recognition: vehicle recognition; smoke recognition, flame recognition; static sign recognition, fire extinguisher recognition, battery vehicle recognition, etc. 4. Environmental change feature recognition: monitoring of floating debris on the water surface, accumulation of debris, camera obstruction; fire passage occupation, illegal parking of non-tank trucks, overflowing garbage cans, etc.

[0030] After long-term independent research and development, and with focus and research on the product, the Xinhuo algorithm has maintained an accuracy rate of over 95% and a false alarm rate of less than 5%, ensuring extremely high user satisfaction. In terms of rapid algorithm customization and low false alarm rate, it fully meets customers' requirements for rapid technical support and core recognition capabilities.

[0031] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0032] Example 1, reference Figure 1-Figure 8, a smart camera provided by an embodiment of the present invention includes a housing 12 and a mainboard 10, wherein the mainboard 10 is installed inside the housing 12 and a communication module 8 is also installed inside the housing 12, a data connection hole is arranged outside the housing 12, a speaker 7, a lens 1, an infrared lamp 2 and a fill light 3 are arranged on the housing 12, the mainboard 10 is electrically connected to the communication module 8, the speaker 7, the lens 1, the infrared lamp 2 and the fill light 3 respectively, the mainboard 10 adopts the Xinhuo algorithm, the Xinhuo algorithm adopts the XHModelv4 model, and the XHModelv4 model includes an image extraction layer, an aggregation feature layer and a classification and positioning layer. The image extraction layer aims to improve the feature extraction capability and reduce the amount of calculation, and its basic constituent unit is a C2f module. Through effective structural design, redundant parameters are reduced and the calculation efficiency is improved, and it has less parameters and excellent feature extraction capability. The image extraction layer includes four parts, the first part includes two Conv modules and a C2f module, the second part and the third part both include a Conv module and a C2f module, and the fourth part includes a Conv module, a C2f module and an SPPF module. The Conv module is a standard convolution module, including a Conv2d layer, a BatchNorm2d layer, and a SiLU activation function. The C2f module includes an input unit, a feature extraction unit, a feature separation unit, a branch processing unit, and a feature merging unit. The input unit is used to receive the feature map from the previous layer. The feature extraction unit extracts the basic features of the input image through the initial standard convolution module. The feature separation unit divides the feature map into two parts (for example, the left half and the right half). The SPPF module includes an input feature map unit, a preliminary feature extraction unit, a pyramid pooling unit, and a convolution unit. The input feature map unit is used to receive the feature map from the previous layer. The preliminary feature extraction unit performs 1x1Conv convolution processing. The pyramid pooling unit has the following functions: use the maximum pooling (MaxPool2d) to operate; use different pooling sizes (1x1, 5x5, 9x9) to operate on the same input feature map. This multi-scale pooling can capture features at different levels; splice the output feature maps of each pooling operation. The convolution unit further processes the concatenated output feature map using a 1x1 Conv convolution layer and outputs the feature map.

[0033] Example 2, reference Figure 1-Figure 8, a smart camera provided by an embodiment of the present invention includes a housing 12 and a mainboard 10, the mainboard 10 is installed inside the housing 12 and a communication module 8 is also installed inside the housing 12, a data connection hole is arranged outside the housing 12, a speaker 7, a lens 1, an infrared lamp 2 and a fill light 3 are arranged on the housing 12, the mainboard 10 is electrically connected to the communication module 8, the speaker 7, the lens 1, the infrared lamp 2 and the fill light 3 respectively, the mainboard 10 adopts the Xinhuo algorithm, the Xinhuo algorithm adopts the XHModelv4 model, and the XHModelv4 model includes an image extraction layer, an aggregation feature layer and a classification positioning layer. The lens 1 adopts an ultra-clear lens of 8 million pixels, which can clearly and accurately capture the required video, the infrared lamp 2 is equipped with an ICR infrared filter type, and a clear image can be obtained at night, the fill light 3 can perform backlight compensation, and also has the effect of preventing fill light overexposure, the fill light 3 includes a white light fill light and an infrared fill light, which are suitable for different situations, and this product can also suppress strong light and perform 3D digital noise reduction on the captured image to further ensure the clarity of the image. The classification and positioning layer includes the prediction layer (Detect), which is responsible for outputting the final detection results, including three outputs: bounding box (x, y, w, h), confidence, and probability of each category; the classification loss uses BCE (binary cross entropy), and the bounding box regression loss uses CIOU.

[0034] Example 3, reference Figure 1-Figure 8 , a smart camera provided by an embodiment of the present invention includes a housing 12 and a mainboard 10, wherein the mainboard 10 is installed inside the housing 12 and a communication module 8 is also installed inside the housing 12, a data connection hole is arranged outside the housing 12, a speaker 7, a lens 1, an infrared lamp 2 and a fill light 3 are arranged on the housing 12, the mainboard 10 is electrically connected to the communication module 8, the speaker 7, the lens 1, the infrared lamp 2 and the fill light 3 respectively, the mainboard 10 adopts the Xinhuo algorithm, the Xinhuo algorithm adopts the XHModelv4 model, and the XHModelv4 model includes an image extraction layer, an aggregation feature layer and a classification and positioning layer. The data connection hole includes a WIFI hole 4, a 4G hole 5 and a tail line hole 6, which can connect WIFI signals and 4G signals, and is suitable for different working conditions. The mainboard 10 is electrically connected to the communication module 8, the speaker 7, the lens 1, the infrared lamp 2 and the fill light 3 respectively through the wire 9, and has good stability in use.

[0035] Example 4, reference Figure 1-Figure 8A smart camera provided by an embodiment of the present invention includes a housing 12 and a mainboard 10. The mainboard 10 is installed inside the housing 12 and a communication module 8 is also installed inside the housing 12. A data connection hole is provided outside the housing 12. A speaker 7, a lens 1, an infrared lamp 2 and a fill light 3 are provided on the housing 12. The mainboard 10 is electrically connected to the communication module 8, the speaker 7, the lens 1, the infrared lamp 2 and the fill light 3 respectively. The mainboard 10 adopts the Xinhuo algorithm, and the Xinhuo algorithm adopts the XHModelv4 model. The XHModelv4 model includes an image extraction layer, an aggregation feature layer and a classification and positioning layer. A heat sink 11 is also connected to the mainboard 10 to dissipate heat for the mainboard 10 at all times to ensure that the whole can work stably for a long time. The communication module 8 adopts a Bluetooth device to achieve fast data exchange.

[0036] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A smart camera, comprising a housing and a mainboard, wherein the mainboard is installed inside the housing and a communication module is also installed inside the housing, a data connection hole is arranged outside the housing, a speaker, a lens, an infrared lamp and a fill light are arranged on the housing, and the mainboard is electrically connected to the communication module, the speaker, the lens, the infrared lamp and the fill light respectively, characterized in that: The main board adopts the Xinhuo algorithm, and the Xinhuo algorithm adopts the XHModelv4 model. The XHModelv4 model includes an image extraction layer, an aggregation feature layer, and a classification and positioning layer.

2. The smart camera according to claim 1, characterized in that: The image extraction layer includes four parts. The first part includes two Conv modules and one C2f module. The second part and the third part both include one Conv module and one C2f module. The fourth part includes one Conv module, one C2f module and one SPPF module.

3. The smart camera according to claim 1 or 2, characterized in that: The classification and positioning layer includes a detection head and a classification head. The detection head is responsible for predicting the bounding box regression value of each anchor box and the confidence of the target existence; the classification head is used to classify each feature map by using average pooling and output the probability distribution of each category.

4. The smart camera according to claim 2, characterized in that: The Conv module is a standard convolution module, including a Conv2d layer, a BatchNorm2d layer and a SiLU activation function.

5. The smart camera according to claim 2 or 4, characterized in that: The C2f module includes an input unit, a feature extraction unit, a feature separation unit, a branch processing unit and a feature merging unit.

6. The smart camera according to claim 2, characterized in that: The SPPF module includes an input feature map unit, a preliminary feature extraction unit, a pyramid pooling unit and a convolution unit.

7. The smart camera according to claim 1, characterized in that: The data connection holes include a WIFI hole, a 4G hole and a tail line hole.

8. The smart camera according to claim 1 or 7, characterized in that: The mainboard is also connected with a heat sink.

Images