CN115013771A - A wisdom street lamp for district monitoring - Google Patents
A wisdom street lamp for district monitoring Download PDFInfo
- Publication number
- CN115013771A CN115013771A CN202210561576.XA CN202210561576A CN115013771A CN 115013771 A CN115013771 A CN 115013771A CN 202210561576 A CN202210561576 A CN 202210561576A CN 115013771 A CN115013771 A CN 115013771A
- Authority
- CN
- China
- Prior art keywords
- frame
- channel
- convolution
- algorithm
- ghost
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000012544 monitoring process Methods 0.000 title claims abstract description 24
- 238000001514 detection method Methods 0.000 claims abstract description 51
- 238000005507 spraying Methods 0.000 claims abstract description 40
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims abstract description 38
- 238000000605 extraction Methods 0.000 claims abstract description 19
- 230000000007 visual effect Effects 0.000 claims abstract description 14
- 239000011159 matrix material Substances 0.000 claims description 30
- 238000004364 calculation method Methods 0.000 claims description 15
- 238000000034 method Methods 0.000 claims description 15
- 238000001914 filtration Methods 0.000 claims description 9
- 230000005284 excitation Effects 0.000 claims description 7
- 230000007246 mechanism Effects 0.000 claims description 7
- 238000013528 artificial neural network Methods 0.000 claims description 5
- 238000011176 pooling Methods 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 230000000694 effects Effects 0.000 claims description 4
- 230000002411 adverse Effects 0.000 claims description 3
- 238000012790 confirmation Methods 0.000 claims description 3
- 230000006872 improvement Effects 0.000 claims description 2
- 238000011835 investigation Methods 0.000 abstract description 5
- 238000010586 diagram Methods 0.000 description 3
- 238000005286 illumination Methods 0.000 description 3
- 238000002347 injection Methods 0.000 description 2
- 239000007924 injection Substances 0.000 description 2
- 239000000243 solution Substances 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 229910002056 binary alloy Inorganic materials 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000009432 framing Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 239000007921 spray Substances 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
- 238000003466 welding Methods 0.000 description 1
Images
Classifications
-
- F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
- F21—LIGHTING
- F21S—NON-PORTABLE LIGHTING DEVICES; SYSTEMS THEREOF; VEHICLE LIGHTING DEVICES SPECIALLY ADAPTED FOR VEHICLE EXTERIORS
- F21S9/00—Lighting devices with a built-in power supply; Systems employing lighting devices with a built-in power supply
- F21S9/02—Lighting devices with a built-in power supply; Systems employing lighting devices with a built-in power supply the power supply being a battery or accumulator
- F21S9/03—Lighting devices with a built-in power supply; Systems employing lighting devices with a built-in power supply the power supply being a battery or accumulator rechargeable by exposure to light
- F21S9/035—Lighting devices with a built-in power supply; Systems employing lighting devices with a built-in power supply the power supply being a battery or accumulator rechargeable by exposure to light the solar unit being integrated within the support for the lighting unit, e.g. within or on a pole
-
- F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
- F21—LIGHTING
- F21V—FUNCTIONAL FEATURES OR DETAILS OF LIGHTING DEVICES OR SYSTEMS THEREOF; STRUCTURAL COMBINATIONS OF LIGHTING DEVICES WITH OTHER ARTICLES, NOT OTHERWISE PROVIDED FOR
- F21V23/00—Arrangement of electric circuit elements in or on lighting devices
- F21V23/04—Arrangement of electric circuit elements in or on lighting devices the elements being switches
- F21V23/0442—Arrangement of electric circuit elements in or on lighting devices the elements being switches activated by means of a sensor, e.g. motion or photodetectors
- F21V23/0471—Arrangement of electric circuit elements in or on lighting devices the elements being switches activated by means of a sensor, e.g. motion or photodetectors the sensor detecting the proximity, the presence or the movement of an object or a person
-
- F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
- F21—LIGHTING
- F21V—FUNCTIONAL FEATURES OR DETAILS OF LIGHTING DEVICES OR SYSTEMS THEREOF; STRUCTURAL COMBINATIONS OF LIGHTING DEVICES WITH OTHER ARTICLES, NOT OTHERWISE PROVIDED FOR
- F21V33/00—Structural combinations of lighting devices with other articles, not otherwise provided for
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/277—Analysis of motion involving stochastic approaches, e.g. using Kalman filters
-
- F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
- F21—LIGHTING
- F21W—INDEXING SCHEME ASSOCIATED WITH SUBCLASSES F21K, F21L, F21S and F21V, RELATING TO USES OR APPLICATIONS OF LIGHTING DEVICES OR SYSTEMS
- F21W2131/00—Use or application of lighting devices or systems not provided for in codes F21W2102/00-F21W2121/00
- F21W2131/10—Outdoor lighting
- F21W2131/103—Outdoor lighting of streets or roads
-
- F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
- F21—LIGHTING
- F21Y—INDEXING SCHEME ASSOCIATED WITH SUBCLASSES F21K, F21L, F21S and F21V, RELATING TO THE FORM OR THE KIND OF THE LIGHT SOURCES OR OF THE COLOUR OF THE LIGHT EMITTED
- F21Y2115/00—Light-generating elements of semiconductor light sources
- F21Y2115/10—Light-emitting diodes [LED]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30196—Human being; Person
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02B—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO BUILDINGS, e.g. HOUSING, HOUSE APPLIANCES OR RELATED END-USER APPLICATIONS
- Y02B20/00—Energy efficient lighting technologies, e.g. halogen lamps or gas discharge lamps
- Y02B20/72—Energy efficient lighting technologies, e.g. halogen lamps or gas discharge lamps in street lighting
Landscapes
- Engineering & Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Sustainable Development (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to the technical field of street lamp equipment, in particular to an intelligent street lamp for community monitoring, which comprises a lamp post, wherein the left side of the top of the lamp post is provided with an upper visual sensor and a lower visual sensor through a bracket, the right side of the top of the lamp post is provided with a solar panel, a photosensitive module and an energy-saving LED lamp through the bracket, the upper end part of the lamp post is provided with a billboard through the bracket, the lower end part of the lamp post is connected with a flower bed water spraying device, the bottom of the lamp post is provided with a box body, and the box body is provided with a road water spraying device. The upper vision sensor has the functions of a foreground target extraction algorithm and a descending target tracking algorithm, so that the function of early warning and the follow-up assistance of staff for responsibility investigation are achieved; the lower vision sensor has the pedestrian detection algorithm and the pressure control device control function, and the pedestrian monitoring and water spraying requirements are met. The solar energy is used for supplying power, so that the energy is saved and the environment is protected; and combine road water jet equipment and flower bed water jet equipment to replace the watering lorry, use manpower resources sparingly.
Description
Technical Field
The invention relates to the technical field of street lamp equipment, in particular to an intelligent street lamp for community monitoring.
Background
Street lamps, which are lamps providing a road with a lighting function, are widely used in various places requiring lighting. The street lamp is the main body of road lighting and also an important carrier of urban lighting function. With the popularization and development of smart cities, the requirements on street lamps are higher and higher, and more functions need to be undertaken.
However, most of the street lamps on both sides of the current residential quarter road only have the functions of illumination, solar power supply and the like, and have larger floor area. In addition, accidents caused by high-altitude throwing are rare, the labor cost is increased continuously, and more manpower is needed for spraying water to the flower bed and the road. Under the big background of wisdom city construction, urgent need a wisdom street lamp.
The current chinese patent CN215908966U discloses an intelligent street lamp, which specifically comprises: when the pedestrian is close to the street lamp, can sense pedestrian's being close through distance sensor and the infrared ray sensor that sets up, at this moment little the control unit can control power control module, and power control module controls the LED lamp plate afterwards and opens, at this moment can realize the effect of automatic switching on, and walk far back at the pedestrian, and power control module can control the LED lamp plate and close, and then the energy can be saved. This publication has the following drawbacks:
1. the distance sensor and the infrared sensor can only be used for simply distinguishing black and white on the road surface, the detection distance and the detection precision are limited, and the detection may be interfered by the difference of the height positions of the sensors.
2. The street lamp function is too few, does not really realize intelligent street lamp, and its each module can not freely dismantle, causes the waste of resource.
Therefore, it is necessary to design an intelligent street lamp for monitoring a cell to solve the above technical problems.
Disclosure of Invention
The invention aims to solve the defects in the prior art, and provides an intelligent street lamp for cell monitoring.A vision sensor has the functions of a foreground target extraction algorithm and a descending target tracking algorithm, so that the function of early warning and the follow-up assistance of staff for responsibility investigation are achieved; the lower vision sensor has the pedestrian detection algorithm and the pressure control device control function, and the pedestrian monitoring and water spraying requirements are met.
In order to achieve the purpose, the invention adopts the following technical scheme:
a smart street lamp for monitoring a community comprises a lamp post, wherein an upper visual sensor and a lower visual sensor are mounted on the left side of the top of the lamp post through a support, a solar panel, a photosensitive module and an energy-saving LED lamp are mounted on the right side of the top of the lamp post through a support, a billboard is mounted on the upper end of the lamp post through a support, a flower bed water spraying device is connected to the lower end of the lamp post, a box body is mounted at the bottom of the lamp post, and a road water spraying device is arranged on the box body;
the upper visual sensor has the functions of a foreground target extraction algorithm and a descending target tracking algorithm, the high-altitude thrown object is locked based on the foreground target extraction algorithm, the high-altitude thrown object is tracked through the descending target tracking algorithm, and a resident with a high-altitude throwing object is reversely searched, so that the function of early warning is achieved, and the follow-up assistance of workers is realized for responsibility investigation;
the lower vision sensor is used for detecting pedestrians passing through the vision area, a lightweight target detection algorithm PD-SSD is used as a core, the pedestrians in the vision area are detected and selected in a frame mode, once the pedestrians enter the spraying area, water spraying is automatically stopped, and adverse effects caused by mistakenly spraying the pedestrians during spraying operation are avoided.
Preferably, the foreground object extraction algorithm takes a three-channel hash algorithm as a core, continuously divides the picture by a quartering method for iterative detection, measures the three-channel hash value of the current frame image R, G, B and the three-channel hash value of the background image respectively by the three-channel hash algorithm, obtains the hash values in a binary form after processing, and finally finds out a set of pixel points of which the hamming distances of the three-channel hash values are smaller than a threshold lambda as a foreground region, namely a high-altitude tossed object;
dividing the collected image information into RGB three channels for processing respectively, defining its background image matrix as H n The current frame image matrix is M n Where n represents n iterations of the quartering method, and the three-channel image matrices of the background frame R, G, B are respectively The image matrixes of three channels of the current frame R, G, B are respectively q represents the number of pixel points in the current area, and the specific operation steps are as follows:
when n is equal to 0, the three-channel image matrix of the background frame R, G, B isThe image matrixes of the three channels of the current frame R, G, B are respectively Comparing R, G, B Hamming distances of hash values of the three channel background frames and the current frame respectively, and if the Hamming distances are smaller than a threshold lambda, determining that the current frame is consistent with the background frame and no high-altitude parabola appears;
if the Hamming distance of one channel is larger than the threshold lambda, the current frame is considered to be inconsistent with the background frame, high-altitude parabolic motion occurs, the n +1 th time is carried out, and 4 th judgment is further carried out respectively n And (3) the Hamming distances of the hash values of three channels of the current frame and the background frame R, G, B in the small region, and finally locking the set of pixel points of which the Hamming distances of the hash values of the three channels are all smaller than a certain threshold value as a foreground region by loop iteration, namely, the high-altitude falling object.
Preferably, the three-channel hash algorithm is an algorithm for comparing the similarity of pictures improved based on the perceptual hash algorithm, that is, the picture information is divided into three channels to be processed and converted into a binary hash value, and whether the two pictures are similar or not is judged by comparing hamming distances of the hash values respectively obtained by the three channels of the two pictures. Taking R channel as an example, defining the image matrix of R channel of background frame as H R ={r′ 1 ,r′ 2 ,r′ 3 ,......r′ q An image matrix of a current frame R channel is M R ={r 1 ,r 2 ,r 3 ,......r q Firstly, two pictures are abbreviated to be 8 multiplied by 8, and the information of high frequency domain in the pictures is removed to obtain a matrixAndrespectively corresponding 64 pixel matrixes of a background frame and a current frame; then, a matrix is obtainedAndaverage values of the elements in the formula are avg H And avg M (ii) a Define the hash value of R channel image of background frame asThe hash value of the R channel image of the current frame isMatrix ofEach element and avg H Comparing, if the element value is greater than or equal to avg H The element value becomes 1, otherwise, the element value becomes 0, finally, the elements in the obtained matrix with only 0 or 1 are combined into a 64-bit integer from the first element, and the integer is stored in a 2-system form, namely, the integer isThe calculation method of (2) is the same as above; final comparisonAndif the Hamming distance between the two channels is smaller than the threshold lambda, the current frame and the R channel image of the background frame are considered to be consistent.
Preferably, the falling target tracking algorithm is used for tracking a high-altitude falling object and reversely searching a high-altitude parabolic householder, and the probability of secondary parabolic movement is greater than that of other users due to the existence of the high-altitude parabolic movement of the householder. The accuracy of a first frame detection result can be effectively enhanced, the Kalman filtering motion variable initialization result is optimized, and the method comprises the following specific steps:
step 1: establishing a corresponding track (Tracks) of a result detected in the first frame, initializing a motion variable of Kalman filtering, introducing a weight adjustment model, accurately predicting a corresponding frame by combining with the Kalman filtering, and selecting a position of a falling object;
step 2: matching the detection frame of the current frame target with the prediction frame of the previous frame orbit (Tracks) and calculating the cost matrix of the detection frame;
step 3: performing cascade matching based on the Hungarian algorithm, if the detection box is successfully matched with the prediction box, successfully tracing, and defining the prediction box of the Tracks (Tracks) as a confirmation state; defining the prediction box of the track (Tracks) as an uncertain state if the prediction boxes are mismatched; if the detection box is mismatched, initializing the detection box as a new track (Tracks) prediction box;
step 4: and repeatedly executing Step2 and Step3 until the tracking is completed or the detection is finished.
Step 5: since the high-altitude parabolic position selected by the first frame of Step1 may be among several households, the responsibility investigation is inconvenient, and the household with the highest weight is selected for box selection to lock the household with the high-altitude parabolic position (the user with the high-altitude parabolic position is selected for box selection).
Preferably, the PD-SSD algorithm comprises: a PDblock network module is built by taking a ghost module as a core and combining an SE (stress independent) attention mechanism, and Resblock in CPSDarknet53 is improved into PDblock; the improved CPSDarknet53 is adopted to replace a backbone network VGG of the SSD.
Preferably, the SE attention module is used to increase the attention of the network model, ignore unimportant features, focus too much on important features, and adjust the degree of importance between channels. The SE attention module comprises squeezing (Squeeze) and Excitation (Excitation), wherein the squeezing operation is firstly carried out, the feature maps are integrated in the spatial dimension, and a channel descriptor is generated to enable a lower network to feel information from the whole world; then, excitation operation is carried out, and the dependence degree among all channels is judged; the specific operation method comprises the following steps:
firstly, extracting features through convolution operation and global operation, and then obtaining a global feature value through a pooling layer, wherein the calculation formula is as follows:
where se (X) represents the global feature value, W, H represents the width and height of the input feature map, and X represents the output after the convolution operation.
Preferably, the ghost module can effectively reduce the size of the neural network and accelerate the model reasoning speed, a ghost operation which is lighter than the convolution operation is used for generating a redundancy characteristic, and the ghost operation is improved relative to the convolution operation; here, the ghost module replaces the whole convolution operation by combining a small number of convolutions with lightweight redundant feature generation, and the specific operations are as follows:
step 1: defining the number of convolution kernels used in the original convolution operation as p, defining the number of convolution kernels used in the ghost operation as q, wherein q is less than p, and generating q characteristic graphs by using q convolution kernels;
step 2: performing deep convolution operation on q feature maps generated by Step1, generating k new feature maps for each feature map, and totaling k × q new feature maps, wherein k × q is p;
step 3: splicing the feature maps together, and replacing the whole convolution by a small amount of convolution and lightweight redundant features; defining the calculation amount of the whole ghost operation as GS, and the calculation amount of the conventional convolution operation as G, then calculating the formula as follows:
G=(p×h′×w′×c×t×t)
GS=(q×h′×w′×c×t×t)+[(k-1)×q×h′×w′×u×u]
wherein G is the calculated amount of convolution operation, GS is the calculated amount of ghost operation, p is the output dimension of convolution operation, q is the output dimension of ghost operation, c is the number of channels, h 'is the height of output, w' is the width of output, t is the height and width of convolution kernel in convolution operation, u is the height and width of convolution kernel in ghost operation, and k is the number of feature maps generated by deep convolution operation; defining the ratio of the calculated amount of the convolution operation and the calculated amount of the ghost operation as F, and calculating the formula as follows:
in the formula, F is a ratio of calculated quantities of product operation and ghost operation, G is a calculated quantity of convolution operation, GS is a calculated quantity of the ghost operation, p is an output dimension of the convolution operation, q is an output dimension of the ghost operation, c is a channel number, h 'is output height, w' is output width, t is height and width of a convolution kernel in the convolution operation, u is height and width of the convolution kernel in the ghost operation, k is the number of feature maps generated by deep convolution operation, k is a positive integer, the calculated quantity of the convolution operation is k times of a ghost module, and the ghost operation is more lightweight, so that the size of a neural network can be effectively reduced, and the model reasoning speed is increased.
Preferably, Resblock in CPSDarknet53 is modified into PDblock, after a convolution, a residual operation is used, part of the input is subjected to feature extraction by a ghost module and an SE attention mechanism module, and the other part of the input is subjected to a convolution operation, and then connected with the extracted features through shortcut to be finally output. The detection accuracy of the CPSDarknet53 can be effectively enhanced, the size of the model and the detection speed are hardly influenced, the improved CPSDarknet53 is adopted to replace a backbone network VGG of the SSD, the last pooling layer, the full-link layer and the Softmax layer are deleted, and the extracted features are directly input into a prediction feature map module of the SSD through a convolution operation; the calculation amount can be effectively reduced, and the model reasoning is accelerated. Compared with the original model, the detection precision is almost unchanged, and the detection speed is greatly improved, so that the actual requirements of the pedestrian detection scene are basically met.
Preferably, the road water spraying device comprises a lifting device and a high-pressure spraying water gun installed on the lifting device, and a control module is arranged on one side of the lifting device; the high-pressure injection water gun can be lifted up and down through the lifting device, the flower bed water spraying device comprises a pressure control module and a rotating block arranged on the pressure control module, and the rotating block is provided with a limit baffle and a high-pressure nozzle. When the flower bed water spraying device works, two high-pressure nozzles on the rotating block are respectively adjusted through the pressure control module to spray water at different water pressures so as to achieve the purposes of rotating and changing the water spraying direction and height, a lower vision sensor on the support is used for detecting pedestrians passing through the vision area, a lightweight target detection algorithm PD-SSD is used as a core, the pedestrians in the vision area are detected and framed, the water spraying is automatically stopped once the pedestrians enter the spraying area, and the bad influence caused by mistakenly spraying the pedestrians during the spraying operation is avoided.
The invention has the following beneficial effects:
1. the solar energy power supply device can utilize solar energy to supply power, and is energy-saving and environment-friendly; and combine road water jet equipment and flower bed water jet equipment to replace the watering lorry, use manpower resources sparingly.
2. According to the invention, the pedestrian monitoring and water spraying requirements are realized through the pedestrian detection algorithm and the pressure control device control function of the lower visual sensor.
3. The invention provides the PD-SSD algorithm aiming at the improvement of the SSD algorithm in the pedestrian monitoring environment, and compared with the SSD algorithm, the PD-SSD algorithm has the advantages of higher detection speed, higher detection precision and stronger robustness in the pedestrian monitoring task.
4. The upper visual sensor is adopted to have a foreground target extraction algorithm and a descending target tracking algorithm, the foreground target extraction algorithm takes a three-channel hash algorithm as a core, a quartering method is adopted to continuously divide and iteratively detect the picture, the three-channel hash value of the current frame image R, G, B and the three-channel hash value of the background image are respectively measured through the three-channel hash algorithm, the hash values obtained after processing are in a binary form, and finally a set of pixel points of which the Hamming distances of the three-channel hash values are smaller than a threshold lambda is found out to be a foreground region, namely an object thrown aloft; and tracking the high-altitude falling object and drawing a falling track by a falling target tracking algorithm, thereby achieving the function of early warning and assisting the following staff to carry out responsibility exploration.
5. The method adopts a descending target tracking algorithm, a resident with a high-altitude parabolic behavior is introduced on the basis of Deep-Sort, the probability of secondary parabolic behavior is greater than other user ideas, and when high-altitude parabolic behavior exists once, an initial frame region of a descending object is positioned and the position of the high-altitude parabolic resident is recorded, so that the weight of the initial frame region and the position of the high-altitude parabolic resident is enhanced, and the initial frame region and the position of the high-altitude parabolic resident are used for accurately positioning the initial frame position of the next high-altitude parabolic behavior and finding the position of the high-altitude parabolic resident; the accuracy of the first frame detection result can be effectively enhanced, and the Kalman filtering motion variable initialization result is optimized.
6. Various modules in the invention can be disassembled, and can be used in various practical scenes, thereby saving resources while meeting the requirements.
Drawings
FIG. 1 is a schematic view of the overall structure of the present invention;
FIG. 2 is a schematic view of the road sprinkler of the present invention;
FIG. 3 is a schematic structural view of the flower bed water spraying device of the present invention;
FIG. 4 is a schematic diagram of a foreground object extraction algorithm in accordance with the present invention;
FIG. 5 is an overall flow chart of a descending target tracking algorithm of the present invention;
FIG. 6 is a schematic diagram of the PD-SSD algorithm of the present invention;
fig. 7 is a schematic structural diagram of the PDBlock principle of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments.
Referring to fig. 1-7, an wisdom street lamp for district monitoring, includes lamp pole 2, there are visual sensor 1-1 and lower visual sensor 1-2 in the top left side of lamp pole 2 through the support mounting, there are solar panel 3, photosensitive module 4 and energy-saving LED lamp 5 on the top right side of lamp pole 2 through the support mounting, there is bill-board 9 in the upper end of lamp pole 2 through the support mounting, the lower tip of lamp pole 2 is connected with flower bed water jet equipment 8, box 6 is installed to the bottom of lamp pole 2, be equipped with road water jet equipment 7 on the box 6.
Wherein, lamp pole 2 is long cartridge type structure, stays enough safe thickness, makes inside cavity make things convenient for access such as cable water pipe, the top calotte, elegant appearance, and side welding has the support for extend the function, the bottom is a box 6, is used for storing road water jet equipment 7's subassembly, and the lower most is the mounting hole, is used for installing this wisdom street lamp.
Wherein, install solar panel 3 in 2 both sides in the lamp pole for increase illumination time and illumination area, acquire more electric energy. The energy-saving LED lamp 5 and the photosensitive module 4 are installed on the lamp post support, and the LED lamp is switched on and off at proper time through the photosensitive module 4.
As shown in fig. 2 to 3, the housing 6 stores therein components of the road sprinkler and an edge calculating box. The road water spraying device 7 comprises a high-pressure spraying water gun 7-1, a lifting device 7-2 and a control module 7-3; the high-pressure injection water gun 7-1 can be lifted up and down through the lifting device 7-2; the flower bed water spraying device 8 comprises a limiting baffle 8-1, a rotating block 8-2, a high-pressure nozzle 8-3 and a pressure control module 8-4.
The upper visual sensor 1-1 has the functions of a foreground target extraction algorithm and a descending target tracking algorithm, and is used for locking an object thrown from high altitude based on the foreground target extraction algorithm, tracking the object thrown from high altitude through the descending target tracking algorithm, reversely searching a resident thrown from high altitude, and achieving the function of early warning and the follow-up assistance of staff for responsibility exploration.
The lower vision sensor 1-2 is used for detecting pedestrians passing through the vision area, a lightweight target detection algorithm PD-SSD is used as a core, the pedestrians in the vision area are detected and framed, once the pedestrians enter the spraying area, water spraying is automatically stopped, and adverse effects caused by mistaken spraying of the pedestrians during spraying operation are avoided.
Specifically, a foreground object extraction algorithm is adopted, a three-channel hash algorithm is taken as a core, a quartering method is adopted to continuously divide the picture for iterative detection, the three-channel hash value of the current frame image R, G, B and the three-channel hash value of the background image are respectively measured through the three-channel hash algorithm, the processed hash values are in a binary form, and finally a set of pixel points of which the Hamming distances of the three-channel hash values are smaller than a threshold lambda is found out to be a foreground areaI.e. objects thrown aloft. Because the high-altitude parabola has certain dangerous factors, the experiment adopts a mode of throwing the bottle from the high altitude to simulate the high-altitude parabola to test the foreground target extraction algorithm on the premise of safety. Fig. 4 is a visualization process of a foreground object extraction algorithm, a three-channel hash algorithm is taken as a core, a four-division method is adopted to continuously divide a picture for iterative detection, three-channel hash values of a current frame image R, G, B and three-channel hash values of a background image are respectively measured through the three-channel hash algorithm, the processed hash values are in a binary form, and finally a set of pixel points with hamming distances of the three-channel hash values smaller than a threshold lambda is found out as a foreground region, namely an object thrown aloft. The three-channel Hash algorithm is an algorithm for comparing the similarity of pictures, which is improved based on the perceptual Hash algorithm, namely, picture information is divided into three channels to be processed and converted into Hash values in a binary system form, and whether two pictures are similar or not is judged by comparing Hamming distances of the Hash values obtained by comparing three channels of the two pictures. Taking R channel as an example, defining the image matrix of R channel of background frame as H R ={r′ 1 ,r′ 2 ,r′ 3 ,......r′ q An image matrix of a current frame R channel is M R ={r 1 ,r 2 ,r 3 ,......r q Firstly, two pictures are abbreviated to be 8 multiplied by 8, and the information of high frequency domain in the pictures is removed to obtain a matrixAndrespectively corresponding 64 pixel matrixes of a background frame and a current frame; then, a matrix is obtainedAndaverage values of the elements in the formula are avg H And avg M (ii) a Defining the hash value of R channel image of background frame asThe hash value of the R channel image of the current frame isMatrix ofEach element and avg H Comparing, if the element value is greater than or equal to avg H The element value becomes 1, otherwise, the element value becomes 0, finally, the elements in the obtained matrix with only 0 or 1 are combined into a 64-bit integer from the first element, and the integer is stored in a 2-system form, namely, the integer isThe calculation method of (2) is the same as above; final comparisonAndif the Hamming distance between the two channels is smaller than the threshold lambda, the current frame and the R channel image of the background frame are considered to be consistent.
The specific operation flow is as follows:
dividing the collected image information into RGB three channels for processing respectively, defining its background image matrix as H n The current frame image matrix is M n Where n represents n iterations of the quartering method, and the three-channel image matrices of the background frame R, G, B are respectively The image matrixes of three channels of the current frame R, G, B are respectively q represents the number of the pixel points in the current area, and the specific operation steps are as follows:
when n is equal to 0, the three-channel image matrix of the background frame R, G, B isThe image matrixes of three channels of the current frame R, G, B are respectively Comparing R, G, B Hamming distances of hash values of the three channel background frames and the current frame respectively, and if the Hamming distances are smaller than a threshold lambda, determining that the current frame is consistent with the background frame and no high-altitude parabola appears;
if the Hamming distance of one channel is larger than the threshold lambda, the current frame is considered to be inconsistent with the background frame, high-altitude parabolic motion occurs, the n +1 th time is carried out, and 4 th judgment is further carried out respectively n And (3) the Hamming distances of the hash values of three channels of the current frame and the background frame R, G, B in the small region, and finally locking the set of pixel points of which the Hamming distances of the hash values of the three channels are all smaller than a certain threshold value as a foreground region by loop iteration, namely, the high-altitude falling object.
Specifically, fig. 5 is an overall flow of a descending target tracking algorithm, and because a resident with a high-altitude parabolic behavior has a probability of parabolic again greater than other users, the descending target tracking algorithm provided by the present invention introduces the above idea on the basis of Deep-Sort, positions the initial frame region of the descending object and records the position of the high-altitude parabolic resident every time there is a high-altitude parabolic, and enhances the weights of the initial frame region and the position of the high-altitude parabolic resident, so as to accurately position the initial frame position of the next high-altitude parabolic and find the position of the high-altitude parabolic resident. The accuracy of the first frame detection result can be effectively enhanced, the Kalman filtering motion variable initialization result is optimized, and the specific operation mode is as follows:
step 1: establishing a corresponding track (Tracks) of a result detected in the first frame, initializing a motion variable of Kalman filtering, and simultaneously introducing a weight adjustment model and accurately predicting a corresponding frame (the framed position is a falling object position) by combining with the Kalman filtering;
step 2: matching the detection frame of the current frame target with the prediction frame of the previous frame orbit (Tracks), and calculating a cost matrix of the detection frame;
step 3: performing cascade matching based on the Hungarian algorithm, if the detection box is successfully matched with the prediction box, successfully tracing, and defining the prediction box of the Tracks (Tracks) as a confirmation state; defining the prediction box of the track (Tracks) as an uncertain state if the prediction boxes are mismatched; if the detection box is mismatched, initializing the detection box as a new track (Tracks) prediction box;
step 4: and repeatedly executing Step2 and Step3 until the tracking is completed or the detection is finished.
Step 5: since the high-altitude parabolic position selected by the first frame at Step1 may be among several households, the responsibility investigation is inconvenient, and the household with the highest weight is selected for frame selection, so that the household with the high-altitude parabolic position is locked (the user with the high-altitude parabolic position is selected for frame selection).
Specifically, the method for framing the object region in the lightweight target detection algorithm is improved on the basis of the target detection algorithm SSD, the PD-SSD algorithm is provided, and the detection speed is greatly improved while the detection accuracy is kept unchanged in the pedestrian monitoring scene. The invention modifies the original algorithm, and the PD-SSD algorithm structure is shown in FIG. 6:
(1) a PDblock network module is established by taking a ghost module as a core and combining an SE attention mechanism, and is used for replacing Resblock in CPSDarknet 53;
(2) and replacing a backbone network VGG of the SSD by the improved CPSDarknet 53.
The SE attention module is used for improving the attention of the network model, neglecting unimportant features, paying more attention to important features and adjusting the importance degree among channels. The SE is the Squeeze (Squeeze) and Excitation (Excitation), and first, a Squeeze operation is performed to integrate the feature maps in the spatial dimension, and a channel descriptor is generated to let the lower network feel the information from the global. And then excitation operation is carried out to judge the dependency degree among the channels. The specific operation method comprises the following steps:
firstly, extracting features through convolution operation and global operation, and then obtaining a global feature value through a pooling layer, wherein the calculation formula is as follows:
where se (X) represents the global feature value, W, H represents the width and height of the input feature map, and X represents the output after the convolution operation.
Specifically, the ghost module can effectively reduce the size of the neural network and accelerate the model reasoning speed. A ghost operation which is lighter than the convolution operation is used for generating the redundant feature, the ghost operation is improved relative to the convolution operation, the whole convolution operation is replaced by a mode of combining a small amount of convolution with the lightweight redundant feature generation, and the specific operation is as follows:
step 1: defining the number of convolution kernels used in the original convolution operation as p, defining the number of convolution kernels used in the ghost operation as q, wherein q is less than p, and generating q characteristic graphs by using q convolution kernels;
step 2: performing deep convolution operation on q feature maps generated by Step1, generating k new feature maps for each feature map, and summing up k × q new feature maps, wherein k × q is p;
step 3: the feature maps are stitched together, and a small number of convolutions plus lightweight redundant features replace the entire convolution. Defining the calculation amount of the whole ghost operation as GS, and the calculation amount of the conventional convolution operation as G, then calculating the formula as follows:
G=(p×h′×w′×c×t×t)
GS=(q×h′×w′×c×t×t)+[(k-1)×q×h′×w′×u×u]
wherein G is the calculated amount of convolution operation, GS is the calculated amount of ghost operation, p is the output dimension of convolution operation, q is the output dimension of ghost operation, c is the number of channels, h 'is the height of output, w' is the width of output, t is the height and width of convolution kernel in convolution operation, u is the height and width of convolution kernel in ghost operation, k is the number of feature maps generated by deep convolution operation, the ratio of the calculated amount of convolution operation and ghost operation is defined as F, and the calculation formula is as follows:
in the formula, F is the ratio of the calculated quantity of the product operation and the calculated quantity of the ghost operation, G is the calculated quantity of the convolution operation, GS is the calculated quantity of the ghost operation, p is the output dimensionality of the convolution operation, q is the output dimensionality of the ghost operation, c is the number of channels, h 'is the output height, w' is the output width, t is the height and the width of a convolution kernel in the convolution operation, u is the height and the width of the convolution kernel in the ghost operation, k is the number of feature maps generated by the deep convolution operation, k is a positive integer, the calculated quantity of the convolution operation is k times of a ghost module, the ghost operation is lighter, the size of a neural network can be effectively reduced, and the model reasoning speed is increased.
The invention combines the advantages of the two modules, fuses the SE attention mechanism and the ghost module to build the PDblock network module, and the specific structure is shown in FIG. 7.
Specifically, Resblock in CPSDarknet53 is improved to PDblock, after convolution, residual operation is used, part of input is subjected to feature extraction by a ghost module and an SE attention mechanism module, and the other part of input is subjected to convolution operation, and then connected with the extracted features through shortcut to be finally output. The detection accuracy of the CPSDarknet53 can be effectively enhanced, the size of the model and the detection speed are hardly influenced, the improved CPSDarknet53 is adopted to replace a backbone network VGG of the SSD, the last pooling layer, the full-link layer and the Softmax layer are deleted, and the extracted features are directly input into a prediction feature map module of the SSD through a convolution operation. The calculation amount can be effectively reduced. And accelerating model reasoning. Compared with the original model, the detection precision is almost unchanged, and the detection speed is greatly improved, so that the actual requirements under the pedestrian detection scene are basically met.
And monitoring the pedestrians in the visual area, performing frame selection by using a lightweight target detection algorithm PD-SSD, and stopping spraying when the pedestrians enter an effective spraying area.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art should be considered to be within the technical scope of the present invention, and the technical solutions and the inventive concepts thereof according to the present invention should be equivalent or changed within the scope of the present invention.
Claims (9)
1. A smart street lamp for monitoring a community comprises a lamp post and is characterized in that an upper visual sensor and a lower visual sensor are mounted on the left side of the top of the lamp post through a support, a solar panel, a photosensitive module and an energy-saving LED lamp are mounted on the right side of the top of the lamp post through a support, a billboard is mounted on the upper end of the lamp post through a support, a flower bed water spraying device is connected to the lower end of the lamp post, a box body is mounted at the bottom of the lamp post, and a road water spraying device is arranged on the box body;
the upper visual sensor has the functions of a foreground target extraction algorithm and a descending target tracking algorithm, the high-altitude thrown object is locked based on the foreground target extraction algorithm, the high-altitude thrown object is tracked through the descending target tracking algorithm, and residents with high-altitude throws are reversely searched, so that the function of early warning is achieved, and the responsibility pursuit of subsequent auxiliary workers is realized;
the lower vision sensor is used for detecting pedestrians passing through the vision area, a lightweight target detection algorithm PD-SSD is used as a core, the pedestrians in the vision area are detected and selected in a frame mode, once the pedestrians enter the spraying area, water spraying is automatically stopped, and adverse effects caused by mistakenly spraying the pedestrians during spraying operation are avoided.
2. The intelligent street lamp for monitoring the residential area as claimed in claim 1, wherein the foreground object extraction algorithm takes a three-channel hash algorithm as a core, the image is continuously segmented and iteratively detected by a quartering method, the three-channel hash value of the current frame image R, G, B and the three-channel hash value of the background image are respectively measured by the three-channel hash algorithm, the processed hash values are in a binary form, and finally a set of pixel points with the hamming distances of the three-channel hash values smaller than a threshold lambda is found out as a foreground area, namely an object thrown aloft;
dividing the collected image information into three channels of RGB for processing respectively, defining its background image matrix as H n The current frame image matrix is M n Where n represents n iterations of the quartering method, and the three-channel image matrices of the background frame R, G, B are respectively The image matrixes of three channels of the current frame R, G, B are respectively q represents the number of pixel points in the current area, and the specific operation steps are as follows:
when n is equal to 0, the three-channel image matrix of the background frame R, G, B isThe image matrixes of three channels of the current frame R, G, B are respectively Comparing R, G, B Hamming distances of hash values of the three channel background frames and the current frame respectively, and if the Hamming distances are smaller than a threshold lambda, determining that the current frame is consistent with the background frame and no high-altitude parabola appears;
if the Hamming distance of one channel is greater than the threshold lambda, the current frame is considered to be inconsistent with the background frame, a high-altitude parabola appears, and the nth + is carried out1 time, and further 4 times of judgment n And (3) the Hamming distances of the hash values of three channels of the current frame and the background frame R, G, B in the small region, and finally locking the set of pixel points of which the Hamming distances of the hash values of the three channels are all smaller than a certain threshold value as a foreground region by loop iteration, namely, the high-altitude falling object.
3. The intelligent street lamp for monitoring the residential quarter as claimed in claim 2, wherein the three-channel hash algorithm is an algorithm for comparing the similarity of pictures based on the improvement of the perceptual hash algorithm, i.e. the picture information is divided into three channels to be processed and converted into a binary hash value, and the hamming distance of the hash values obtained by comparing the three channels of the two pictures is used for judging whether the two pictures are similar; taking R channel as an example, defining the image matrix of R channel of background frame as H R ={r′ 1 ,r′ 2 ,r′ 3 ,......r′ q An image matrix of a current frame R channel is M R ={r 1 ,r 2 ,r 3 ,......r q Firstly, two pictures are abbreviated to be 8 multiplied by 8, and the information of high frequency domain in the pictures is removed to obtain a matrixAndrespectively corresponding 64 pixel matrixes of a background frame and a current frame; then, a matrix is obtainedAndaverage values of the elements in the formula are avg H And avg M (ii) a Defining the hash value of R channel image of background frame asThe hash value of the R channel image of the current frame isMatrix arrayEach element and avg H Comparing, if the element value is greater than or equal to avg H The element value is changed into 1, otherwise, the element value is changed into 0, finally, the elements in the obtained matrix with only 0 or 1 are combined into a 64-bit integer from the first element, and the integer is stored in a 2-system form, namely, the integer is the elementThe calculation method of (3) is as above; final comparisonAndif the Hamming distance between the two channels is smaller than the threshold lambda, the current frame and the R channel image of the background frame are considered to be consistent.
4. The intelligent street lamp for monitoring the residential area as claimed in claim 1, wherein the descending target tracking algorithm is used for tracking the high-altitude falling object and reversely searching the high-altitude parabolic householder, and comprises the following steps:
step 1: establishing a corresponding track of a result detected in the first frame, initializing a motion variable of Kalman filtering, introducing a weight adjustment model and accurately predicting a corresponding frame by combining Kalman filtering, wherein the frame is selected as a falling object position;
step 2: matching the detection frame of the current frame target with the prediction frame of the previous frame track, and calculating a cost matrix of the detection frame;
step 3: performing cascade matching based on the Hungarian algorithm, if the detection box is successfully matched with the prediction box, successfully tracing, and defining the prediction box of the track as a confirmation state; if the prediction boxes are mismatched, defining the prediction box of the track as an uncertain state; if the detection box is mismatched, initializing the detection box as a new track prediction box;
step 4: and repeatedly executing Step2 and Step3 until the tracking is completed or the detection is finished.
5. The intelligent street lamp for monitoring cells as claimed in claim 1, wherein the PD-SSD algorithm comprises: a PDblock network module is built by taking a ghost module as a core and combining an SE (stress independent) attention mechanism, and Resblock in CPSDarknet53 is improved into PDblock; and replacing a backbone network VGG of the SSD by the improved CPSDarknet 53.
6. The intelligent street lamp for monitoring the cells as claimed in claim 5, wherein the SE attention module first performs a squeezing operation to integrate the feature map in the spatial dimension to generate a channel descriptor for the lower network to feel the information from the global; then, excitation operation is carried out, and the dependence degree among all channels is judged; the specific operation method comprises the following steps:
firstly, extracting features through convolution operation and global operation, and then obtaining a global feature value through a pooling layer, wherein the calculation formula is as follows:
where se (X) represents the global feature value, W, H represents the width and height of the input feature map, and X represents the output after the convolution operation.
7. The intelligent street lamp for monitoring the cells as claimed in claim 5, wherein the ghost module replaces the whole convolution operation by combining a small number of convolutions with lightweight redundant feature generation, and the operation is as follows:
step 1: defining the number of convolution kernels used in the original convolution operation as p, defining the number of convolution kernels used in the ghost operation as q, wherein q is less than p, and generating q characteristic graphs by using q convolution kernels;
step 2: performing deep convolution operation on q feature maps generated by Step1, generating k new feature maps for each feature map, and summing up k × q new feature maps, wherein k × q is p;
step 3: splicing the feature maps together, and replacing the whole convolution by a small amount of convolution and lightweight redundant features; defining the calculation amount of the whole ghost operation as GS, and the calculation amount of the conventional convolution operation as G, then calculating the formula as follows:
G=(p×h′×w′×c×t×t)
GS=(q×h′×w′×c×t×t)+[(k-1)×q×h′×w′×u×u]
wherein G is the calculated amount of convolution operation, GS is the calculated amount of ghost operation, p is the output dimension of convolution operation, q is the output dimension of ghost operation, c is the number of channels, h 'is the height of output, w' is the width of output, t is the height and width of convolution kernel in convolution operation, u is the height and width of convolution kernel in ghost operation, and k is the number of feature maps generated by deep convolution operation; defining the ratio of the calculated amount of the convolution operation and the calculated amount of the ghost operation as F, and calculating the formula as follows:
in the formula, F is the ratio of the calculated quantity of the product operation and the calculated quantity of the ghost operation, G is the calculated quantity of the convolution operation, GS is the calculated quantity of the ghost operation, p is the output dimensionality of the convolution operation, q is the output dimensionality of the ghost operation, c is the number of channels, h 'is the output height, w' is the output width, t is the height and the width of a convolution kernel in the convolution operation, u is the height and the width of the convolution kernel in the ghost operation, k is the number of feature maps generated by the deep convolution operation, k is a positive integer, the calculated quantity of the convolution operation is k times of a ghost module, the ghost operation is lighter, the size of a neural network can be effectively reduced, and the model reasoning speed is increased.
8. The intelligent street lamp for monitoring the cells as claimed in claim 5, wherein Resblock in CPSDarknet53 is modified into PDblock, after a convolution operation, residual error operation is used, one part of input is subjected to feature extraction by a ghost module and an SE attention mechanism module, and the other part of input is subjected to a convolution operation and then is connected with the extracted features through shortcut for final output.
9. The intelligent street lamp for monitoring the residential quarter as claimed in claim 1, wherein the road water spraying device comprises a lifting device and a high-pressure water spraying gun mounted on the lifting device, and a control module is arranged on one side of the lifting device; the flower bed water spraying device comprises a pressure control module and a rotating block arranged on the pressure control module, and the rotating block is provided with a limiting baffle and a high-pressure nozzle.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210561576.XA CN115013771A (en) | 2022-05-23 | 2022-05-23 | A wisdom street lamp for district monitoring |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210561576.XA CN115013771A (en) | 2022-05-23 | 2022-05-23 | A wisdom street lamp for district monitoring |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115013771A true CN115013771A (en) | 2022-09-06 |
Family
ID=83068405
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210561576.XA Pending CN115013771A (en) | 2022-05-23 | 2022-05-23 | A wisdom street lamp for district monitoring |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115013771A (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160283784A1 (en) * | 2015-03-24 | 2016-09-29 | Michael Kounavis | Multi-layer skin detection and fused hand pose matching |
CN113850367A (en) * | 2021-08-31 | 2021-12-28 | 荣耀终端有限公司 | Network model training method, image processing method and related equipment thereof |
CN113847569A (en) * | 2021-09-30 | 2021-12-28 | 恒明星光智慧文化科技(深圳)有限公司 | Early warning monitoring system and method for high-altitude parabolic intelligent street lamp |
CN114463727A (en) * | 2022-01-11 | 2022-05-10 | 上海工程技术大学 | Subway driver behavior identification method |
-
2022
- 2022-05-23 CN CN202210561576.XA patent/CN115013771A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160283784A1 (en) * | 2015-03-24 | 2016-09-29 | Michael Kounavis | Multi-layer skin detection and fused hand pose matching |
CN113850367A (en) * | 2021-08-31 | 2021-12-28 | 荣耀终端有限公司 | Network model training method, image processing method and related equipment thereof |
CN113847569A (en) * | 2021-09-30 | 2021-12-28 | 恒明星光智慧文化科技(深圳)有限公司 | Early warning monitoring system and method for high-altitude parabolic intelligent street lamp |
CN114463727A (en) * | 2022-01-11 | 2022-05-10 | 上海工程技术大学 | Subway driver behavior identification method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wang et al. | Evolving boxes for fast vehicle detection | |
CN110321853B (en) | Distributed cable external-damage-prevention system based on video intelligent detection | |
Eng et al. | An automatic drowning detection surveillance system for challenging outdoor pool environments | |
CN109416413A (en) | Solar energy forecast | |
CN109635750A (en) | A kind of compound convolutional neural networks images of gestures recognition methods under complex background | |
Kang et al. | Voxel-based extraction and classification of 3-D pole-like objects from mobile LiDAR point cloud data | |
Jia et al. | FoveaMask: A fast and accurate deep learning model for green fruit instance segmentation | |
CN115049944A (en) | Small sample remote sensing image target detection method based on multitask optimization | |
CN105574468A (en) | Video flame detection method, device and system | |
CN105931267A (en) | Moving object detection and tracking method based on improved ViBe algorithm | |
CN103426008A (en) | Vision human hand tracking method and system based on on-line machine learning | |
Wang et al. | Robust and fast vehicle turn-counts at intersections via an integrated solution from detection, tracking and trajectory modeling | |
CN115013771A (en) | A wisdom street lamp for district monitoring | |
Li et al. | Intelligent transportation video tracking technology based on computer and image processing technology | |
Huang et al. | Overview of LiDAR point cloud target detection methods based on deep learning | |
CN103020987A (en) | Quick foreground detection method based on multi-background model | |
Zhao et al. | An Improved Method for Infrared Vehicle and Pedestrian Detection Based on YOLOv5s | |
Jiang et al. | Research progress of single-stage small target detection based on deep learning | |
CN114419356A (en) | Detection method, system, equipment and storage medium for densely-arranged power equipment | |
He et al. | Fast vehicle detection with lateral convolutional neural network | |
Kun et al. | Multi-target detection in airport scene based on Yolov5 | |
JP2000353231A (en) | Picture processor | |
Hong et al. | Improved SSD model for pedestrian detection in natural scene | |
Yu et al. | An Accurate Defect Detection Model for Photosensitive Elements Based on Improved YOLOV4 Deep Neural Network | |
Sun et al. | A YOLOv4-based vehicle detection method from UAV Videos |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |