CN108460336A - A kind of pedestrian detection method based on deep learning - Google Patents
A kind of pedestrian detection method based on deep learning Download PDFInfo
- Publication number
- CN108460336A CN108460336A CN201810082310.0A CN201810082310A CN108460336A CN 108460336 A CN108460336 A CN 108460336A CN 201810082310 A CN201810082310 A CN 201810082310A CN 108460336 A CN108460336 A CN 108460336A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- layers
- deep learning
- candidate person
- pedestrian detection
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
Abstract
The present invention relates to a kind of pedestrian detection methods based on deep learning, and this approach includes the following steps：Video image input extraction network to be detected is generated into characteristic pattern first, then, the characteristic pattern input area that network generates will be extracted and propose network, propose that method detection is most likely to be the region of pedestrian using region, pedestrian candidate person and the corresponding score of pedestrian candidate person are generated, finally determines whether pedestrian candidate person is real pedestrian using trained decision Tree algorithms.It is an advantage of the invention that calculating simply, quickly, the accuracy rate of pedestrian detection can be significantly improved.
Description
Technical field
The present invention relates to a kind of pedestrian detection method, especially a kind of pedestrian detection method based on deep learning belongs to
Technical field of image processing.
Background technology
It is reported that pedestrian detection is an important subject of computer vision field.The purpose of pedestrian detection be for
Accurately identify and position position of the pedestrian in image or video sequence.Currently, pedestrian detection vehicle DAS (Driver Assistant System),
It is widely used in intelligent video monitoring and intelligent transportation.
Traditional pedestrian detection method is also referred to as the model of hand-designed, is the spy that pedestrian is indicated based on low-level image feature
Sign, such as HOG features, Haar features, LBP features, LUV features, ICF features, Squares ChnFtrs features and LDCF are special
Sign etc..At present mostly pedestrian detection is solved the problems, such as grader using support vector machines or decision tree.However, traditional row
People's detection method needs the feature of engineer complex, needs a large amount of professional knowledge and there are one in terms of robustness
Fixed limitation.
With the development of deep learning, based on the pedestrian detection method of deep learning in the case where not considering to calculate cost
Achieve huge success.In general, the pedestrian detection method based on deep learning can be divided into two classes：One kind is to be based on area
The pedestrian detection method that domain is suggested, such as R-CNN methods, SPP-Net methods, Faster R-CNN methods and R-FCN methods；
Another kind of is the pedestrian detection method for being not based on region suggestion, such as YOLO methods and SSD methods.Although being not based on region to build
The pedestrian detection method of view has some advantages in calculating speed, but this method cannot obtain very high precision.Therefore, right
For most of pedestrian detection methods based on deep learning, the plan of pedestrian's candidate is often generated using region suggestion
Slightly.
In addition, as deep learning is in the extensive use of pedestrian's detection field, convolutional neural networks are widely used in
Pedestrian detection, such as AlexNet networks, VGG networks, ZF networks, Fast-RCNN networks, Faster-RCNN networks, R-CNN nets
Network, MS-CNN networks and R-FCN networks.In R-CNN, Fast-RCNN and Faster R-CNN this series of method,
Propose that strategy is used to improve target detection accuracy rate and calculating speed in region.For MS-CNN methods, one multiple dimensioned
Propose that network is used to improve the accuracy of detection Small object in region.For R-FCN methods, carried using full convolutional network and region
View network is combined to carry out pedestrian detection.Compared with Faster-RCNN methods, R-FCN methods substantially increase calculating speed
And slightly improve the accuracy rate of pedestrian detection.Although the target detection technique development based on deep learning is very rapid,
Either in terms of accuracy rate or speed, the method for pedestrian detection is still significantly improved space.
Invention content
It is an object of the invention to：In view of the defects existing in the prior art, a kind of new row based on study depth is proposed
People's detection method, accuracy and rapidity for improving pedestrian detection.
In order to reach object above, the present invention provides a kind of pedestrian detection method based on deep learning, including it is following
Step：
Video image input extraction network to be detected is generated characteristic pattern by the first step；
Second step will extract the characteristic pattern input area proposal network that network generates, and region is recycled to propose that method detection most has
It may be the region of pedestrian, generate pedestrian candidate person and the corresponding score of pedestrian candidate person；
Third step determines whether pedestrian candidate person is real pedestrian using trained decision Tree algorithms.
Video image is input in designed pedestrian detection model by the present invention first, then using based on deep learning
PVANet networks generate characteristic pattern, propose that network generates pedestrian candidate person and corresponding score, last profit followed by region
Classified to the pedestrian candidate person of generation to find out real pedestrian with trained decision Tree algorithms.
Preferably, the extraction network uses PVANet networks, the PVANet networks to have 14 layers, and wherein three first layers are
Convolutional layer, centre are two groups of initial layers, and every group of initial layers include the identical initial layers of four structures, and last three layers are full connection
Layer；The output of the full articulamentum is the input that network and decision tree classifier are proposed in region.
It is further preferred that the structure of all initial layers is all identical, single initial layers are by first, second, third point
Zhi Zucheng, first branch are made of one 1 × 1 convolutional layer, and second branch is by one 1 × 1 convolutional layer and one
A 3 × 3 convolutional layer composition, the third branch are made of one 1 × 1 convolutional layer and two 3 × 3 convolutional layers.
Still further preferably, the specific method is as follows for single initial layers generation characteristic pattern：The characteristic pattern that last layer generates
It is passed to the first, second, third, etc. three branches of initial layers respectively, the characteristic pattern then exported by these three branches is transmitted
To an articulamentum, next layer is finally entered, becomes next layer of input feature vector figure, is more precisely obtained by initial layers in this way
The target of small scale.Last layer is that the next layer of convolutional layer of single initial layers either front is single initial layers or subsequent
Full articulamentum.
Preferably, in network is proposed in the region, for the input feature vector figure that PVANet networks generate, by a m*m
The sliding window of size is used and generates multiple features connected entirely on each width characteristic pattern, and the feature each connected entirely includes two
Branch, one of branch are scs layers, another branch is cds layers.Single sliding window can predict simultaneously different scale and
The region motion of different aspect ratios.For example, when the region motion of sliding window four scales of prediction and four length-width ratios, will produce
4*4 region motion.That is, sliding window generates 4*4*4 output at cds layers, 2*4*4 output is generated at scs layers.
Described cds layers is used for generating pedestrian candidate person, including pedestrian candidate person（Predict target）The coordinate of central point and the pedestrian
Candidate（Predict target）Width and height；Described scs layers is used for generating the corresponding score of pedestrian candidate person, that is, predicts mesh
Target corresponding scores, predict the corresponding scores of target be predict target be the estimated probability of pedestrian and be not pedestrian probability；
The pedestrian candidate person generated by cds layers score corresponding with the pedestrian candidate person generated by scs layers is transported to decision tree classification
Device is trained and detects.
Cost is calculated in order to be reduced in the case where not influencing pedestrian detection precision, in the training process, by PVANet nets
The convolution feature that network generates is used as region and proposes network and detect the input of network.
The decision tree of the present invention uses tree, wherein a characteristic attribute is sentenced in the expression of each non-leaf nodes
It is disconnected, characteristic attribute of each branching representation pair judge as a result, each leaf node represents a classification.In order to determine
Plan, the characteristic attribute for first having to treat class object since root node are tested, and are then selected according to test result corresponding
Branch, finally repeat the process until reaching a leaf node.The classification of the leaf node of arrival is exactly to predict
The classification of the target to be classified.
Although the mutation of decision Tree algorithms has very much, such as ID3, C5.0 and CART algorithms, their basic thought
All it is identical, and the use of the accuracy rate that decision Tree algorithms are classified is very high.Decision tree classifier it is basic
Thought is the multiple Weak Classifiers of training on the same training set, then these Weak Classifiers are combined into final strong classification
Device.These Weak Classifiers are respectively there are one weighting parameter β, that is, the ratio of sample number that grader is correctly classified.Therefore it needs
One threshold value is set to determine whether sample is correctly classified.Successive ignition will be carried out in the training process to decision tree,
If the classification accuracy of some Weak Classifier is very low during an iteration, the performance of the Weak Classifier is also meaned that very
Difference, then the parameter of the Weak Classifier will be reduced.
Specifically, training the method for decision tree as follows using RealBoost algorithms：
1. given training set,
（x1, y1）...（xi, yi）...（xN, yN）
Wherein, yiIt is feature vector, and i=1 ..., N；
2. in the starting stage, Weak Classifier is numbered and number is denoted as j, according to（1）Formula determines each Weak Classifier
Weight,
（1）
Wherein, WjFor the weight of Weak Classifier, H is the number of Weak Classifier；
3. carrying out n times to Weak Classifier to train to obtain training data, training is numbered and number is denoted as n, then basis
（2）Formula obtains a probability Estimation,
（2）
Wherein, Pn（y）For the probabilistic estimated value of Weak Classifier, N is the frequency of training of Weak Classifier；
4. basis（3）Formula calculates the true Distribution value of Weak Classifier,
（3）
Wherein, fn（y）For the actual value of Weak Classifier, R is set of real numbers；
5. in the training process, according to（4）Formula obtains the weight of Weak Classifier,
（4）
6. after each iteration, the weight of all Weak Classifiers is normalized again so that total weight of all Weak Classifiers
Equal to 1, strong classifier is finally obtained, according to（5）Formula obtains strong classifier,
（5）
Wherein, N is the frequency of training of Weak Classifier.
The method for determining real pedestrian is as follows：First with trained decision Tree algorithms to the pedestrian candidate person of generation into
Row classification, then according to pedestrian candidate person's classification preset threshold value, when the pedestrian candidate person in characteristic pattern is the probability of pedestrian
When less than preset threshold value, then pedestrian candidate person classification is real pedestrian, and the otherwise classification is not real pedestrian.
It is an advantage of the invention that calculating simply, quickly, the accuracy rate of pedestrian detection can be significantly improved.
Description of the drawings
The present invention will be further described below with reference to the drawings.
Fig. 1 is the network model figure of the present invention.
Fig. 2 is the structural model figure of PVANet networks in the present invention.
Fig. 3 is the structural model figure of PVANet network initial layers in the present invention.
Fig. 4 is the structural model figure that network is proposed in region in the present invention.
Fig. 5（a）For the part training sample exemplary plot in the present invention on Caltech pedestrian detections data set.
Fig. 5（b）Sample instantiation figure is detected for the part in the present invention on Caltech pedestrian detections data set.
Fig. 6（a）For the part training sample exemplary plot in the present invention on INRIA pedestrian detections data set.
Fig. 6（b）Sample instantiation figure is detected for the part in the present invention on INRIA pedestrian detections data set.
Fig. 7（a）The part sample being detected on Caltech pedestrian detection data sets for pedestrian detection model in the present invention
This exemplary plot.
Fig. 7（b）The part knot being detected on Caltech pedestrian detection data sets for pedestrian detection model in the present invention
Fruit exemplary plot.
Fig. 8（a）The part sample being detected on INRIA pedestrian detection data sets for pedestrian detection model in the present invention
Exemplary plot.
Fig. 8（b）The partial results being detected on INRIA pedestrian detection data sets for pedestrian detection model in the present invention
Exemplary plot.
Specific implementation mode
Embodiment one
A kind of pedestrian detection method based on deep learning is present embodiments provided, the mentality of designing of this method is：By video figure
As being input to designed pedestrian detection model（Model structure is shown in Fig. 1）In, it is given birth to using the PVANet networks based on deep learning
Propose that network proposes that method generates pedestrian candidate person and row using region at characteristic pattern, then by the characteristic pattern input area of generation
The corresponding score of people's candidate finally classifies to find out to the pedestrian candidate person of generation using trained decision Tree algorithms
Real pedestrian.
PVANet networks share 14 layers, and wherein three first layers are convolutional layer, and centre is two groups of initial layers, every group of initial layers packet
Containing the identical initial layers of four structures, last three layers are full articulamentum（It is shown in Table 1）.As shown in Fig. 2, video image enters PVANet
After network, pass through three first layers convolutional layer successively, the processing of intermediate eight layers of initial layers and rear three layers of full articulamentum generates the spy of output
Sign figure, characteristic pattern are passed to region by full articulamentum and propose that network and decision tree classifier, the output of full articulamentum are proposed for region
The input of network and decision tree classifier.The structure of all initial layers is all identical.
The structure for the PVANet networks that table 1. optimized
| Convolutional layer | 4*4_32 |
| Convolutional layer | 3*3_32 |
| Convolutional layer | 3*3_32 |
| First group of initial layers | 1*1_96 - 1*1_16_3*3_64 - 1*1_16_3*3_32_3*3_32 |
| First group of initial layers | 1*1_96 - 1*1_16_3*3_64 - 1*1_16_3*3_32_3*3_32 |
| First group of initial layers | 1*1_96 - 1*1_16_3*3_64 - 1*1_16_3*3_32_3*3_32 |
| First group of initial layers | 1*1_96 - 1*1_16_3*3_64 - 1*1_16_3*3_32_3*3_32 |
| Second group of initial layers | 1*1_128 - 1*1_32_3*3_96 - 1*1_16_3*3_32_3*3_32 |
| Second group of initial layers | 1*1_128 - 1*1_32_3*3_96 - 1*1_16_3*3_32_3*3_32 |
| Second group of initial layers | 1*1_128 - 1*1_32_3*3_96 - 1*1_16_3*3_32_3*3_32 |
| Second group of initial layers | 1*1_128 - 1*1_32_3*3_96 - 1*1_16_3*3_32_3*3_32 |
| Full articulamentum | 4096 |
| Full articulamentum | 4096 |
| Full articulamentum | 1000 |
Table 1 lists the structure of the PVANet networks optimized.In the entire network, l*l_M indicates that the convolution kernel of this layer is l*l
And will export M characteristic pattern.In initial layers, branch different in initial layers is indicated using "-", last full articulamentum
Parameter indicates the number for the neuron for including in full articulamentum.
As shown in figure 3, single initial layers are made of the first, second, third branch, the first branch is by one 1 × 1 convolution
Layer composition, the second branch are made of one 1 × 1 convolutional layer and one 3 × 3 convolutional layer, and third branch is by one 1 × 1
Convolutional layer and two 3 × 3 convolutional layers composition.The characteristic pattern that last layer generates is passed to three branches of initial layers respectively, so
The characteristic pattern exported afterwards by these three branches is transferred into an articulamentum, and the characteristic pattern which exports finally enters next
Layer, becomes next layer of input feature vector figure.Last layer can be the convolutional layer that single initial layers can also be front, and next layer can
To be that single initial layers can also subsequent full articulamentum.The convolution feature generated by PVANet networks is used as region and proposes net
Network and the input for detecting network.
As shown in figure 4, in network is proposed in region, it is for the input feature vector figure that PVANet networks generate, a m*m is big
Small sliding window is used and generates multiple features connected entirely on each width characteristic pattern, and the feature each connected entirely includes two points
Branch, one of branch is scs layers, another branch is cds layers.Single sliding window can predict different scale and not simultaneously
With the region motion of aspect ratio.Cds layers be used for generating the i.e. coordinates of prediction target's center point and the width of the prediction target and
Highly, the scs layers of corresponding scores for being used for generating prediction target predict that target is the estimated probability of pedestrian and is not pedestrian
Probability.The output finally generated by cds layers and by scs layers is sent to decision tree classifier and is trained and detects.
In decision tree classifier, the method for training decision tree is as follows：Given training set（x1, y1）...（xi, yi）...（xN, yN）,, wherein yiIt is feature vector, and i=1 ..., N.In the starting stage, Weak Classifier is numbered simultaneously
Number is denoted as j, according to（1）Formula isDetermine the weight of each Weak Classifier, wherein WjFor Weak Classifier
Weight, H be Weak Classifier number.In the training process, n times first are carried out to Weak Classifier to train to obtain training data, is being instructed
Practice and training is numbered before starting and number is denoted as n, further according to（2）Formula is
Obtain a probability Estimation, wherein Pn（y）For the probabilistic estimated value of Weak Classifier, N is the frequency of training of Weak Classifier, then root
According to（3）Formula isCalculate the true Distribution value of Weak Classifier, wherein fn（y）It is weak
The actual value of grader, R are set of real numbers, last basis（4）Formula isIt obtains weak
The weight of grader.After each iteration, the weight of all Weak Classifiers is normalized again so that all Weak Classifiers
Total weight is equal to 1, finally obtains strong classifier。
After decision tree trains, classified to the pedestrian candidate person of generation using trained decision Tree algorithms, so
Afterwards according to pedestrian candidate person's classification preset threshold value, set in advance when the probability that the pedestrian candidate person in characteristic pattern is pedestrian is less than
When fixed threshold value, then pedestrian candidate person classification is real pedestrian, and the otherwise classification is not real pedestrian.
Pedestrian detection model of the present embodiment based on deep learning is tested, to assess its performance.Specific assessment side
Method is as follows：
Step A, the data set for assessment algorithm performance is introduced.
The experiment uses two pedestrian detection data sets, respectively Caltech pedestrian detections data set and INRIA pedestrian's inspection
Measured data collection.Wherein, Caltech pedestrian detections data set is one bigger pedestrian's data set of current scale, the data
Collection has 250000 frames to be marked from a video for being up to ten hours using vehicle-mounted camera shooting in the video,
It includes 350000 rectangle frames and 2300 different pedestrians that entire video, which has altogether, and the data set is also to these rectangle frames
Between hiding relation marked.This data set includes 11 small videos from entire video, each video
For size all in 1G or so, mark all has been carried out in the first six video therein, this six videos include 192000 pedestrians altogether,
6100 positive samples and 61000 negative samples, are used as being trained pedestrian detection network；Five videos do not correspond to afterwards
Markup information, this five videos altogether include 155000 pedestrians, 56000 positive samples and 65000 negative samples, be used to
Detect the effect of pedestrian detection method.Fig. 5（a）And Fig. 5（b）Give one on the pedestrian detection data set of California Institute of Technology
The example of a little pictures, wherein Fig. 5（a）Show the part training image of Caltech pedestrian detection data sets, Fig. 5（b）Exhibition
What is shown is the partial test image of Caltech pedestrian detection data sets.From Fig. 5（a）And Fig. 5（b）In as can be seen that Caltech
Video frame in pedestrian detection data set is very fuzzy, therefore it is one to carry out experiment on Caltech pedestrian detection data sets
Challenging task.
INRIA pedestrian detection data sets are most common static pedestrian's Test databases, include original graph in the data set
Piece and corresponding label.INRIA data sets provide two kinds of training and test sample, and it is different that one is photo resolutions
Training and test sample, another kind are the identical training of photo resolution and test sample.The experiment of the present embodiment uses picture
The different sample of resolution ratio, wherein training set include 614 positive samples and 1218 negative samples, and test set includes 288 positive samples
Sheet and 453 negative samples.Fig. 6（a）And Fig. 6（b）Give the example of some pictures on INRIA data sets, Fig. 6（a）It is
Train picture, including 6 positive samples and 6 negative samples, Fig. 6 in the part of INIRIA data sets（b）It is the portion of INIRIA data sets
Divide test pictures, including 6 positive samples and 6 negative samples.From Fig. 6（a）And Fig. 6（b）As can be seen that in INRIA data sets
Video definition is relatively high.
Step B, a kind of pedestrian detection method based on deep learning is given for the present embodiment pedestrian detection model.
First, using the characteristic pattern of the mobile pedestrian of PVANet networks extraction optimized（The knot of the PVANet networks of optimization
Structure is as shown in table 1）.Video image can generate 512 characteristic patterns by PVANet networks, preceding 128 features in these characteristic patterns
Figure is used for region and proposes that network generates pedestrian candidate person and corresponding score.
Secondly, propose that network generates pedestrian candidate person and corresponding score using region.Propose to use in network in region
Five kinds of scales and five kinds of length-width ratios generate 25 regions to each sliding window and suggest.For every frame picture, we, which only obtain, divides
Highest 200 regions, which suggest being sent into decision tree classifier, is trained decision tree.
Finally, it trains decision tree classifier and the pedestrian candidate person of generation is divided using trained decision Tree algorithms
Class is to find out real pedestrian.Purpose to decision tree classifier training is succinct and being capable of fine arranged row in order to obtain one
The decision tree of people.In the most initial stage of training decision tree, by all positive samples, grab sample and identical as positive sample quantity
Negative sample and a certain proportion of sample for being difficult to classify as training set, entire training is divided into six stages, by decision
The threshold value of Tree Classifier is set as 0.7.Include 64 in the first stage of training decision tree to set, later the quantity of tree of each stage
It is double, and addition certain proportion, the negative sample for being difficult to classify constitute new training set in old training set, ultimately generate one
The decision tree classifier set with 2048, this decision tree classifier is the strong classifier finally used.
Step C, by contrast experiment, the performance of the put forward algorithm of the present embodiment is inquired into.
During the experiment, the threshold value of decision tree classifier is set as 0.7, to the pedestrian detection model of the present embodiment into
Row experiment.After experiment, for pedestrian detection model experimental result time performance with it is state-of-the-art with some on miss rate
Method is compared, including CompACT-Deep methods, CCF methods and LDCF methods.Final comparison result is in table 2
Middle display.From in final comparison result it can be found that the method for the present embodiment for these state-of-the-art methods not
Advantage is only occupied in the processing of single frames picture and on miss rate again smaller than these methods.
Comparison result of the table 2. on time performance and miss rate
| Method | Time/per pictures (second) | Miss rate % |
| PVANet+RPN+BF [the present embodiment method] | 0.48 | 9 |
| CompACT-Deep | 0.5 | 12 |
| LDCF | 0.6 | 25 |
| CCF | 13 | 17 |
In addition, Fig. 7（a）And Fig. 7（b）Illustrate the sediment that pedestrian detection is carried out on Caltech pedestrian detection data sets
Picture.Fig. 7（a）It is the original image on Caltech data sets, Fig. 7（b）It is that these original images are examined in the pedestrian of the present embodiment
Survey corresponding testing result on model.From Fig. 7（a）And Fig. 7（b）Although can be seen that on Caltech pedestrian detection data sets
Picture is very fuzzy, but the pedestrian detection model of the present embodiment remains to obtain the result of good pedestrian detection.
The pedestrian detection model that the present embodiment is proposed carried out on Caltech pedestrian detection data sets training and
It is tested on INRIA data sets to verify the validity of model.Fig. 8（a）And Fig. 8（b）It illustrates in INRIA pedestrian detection numbers
According to the sediment picture for carrying out pedestrian detection on collection.Fig. 8（a）It is the original image on INRIA data sets, Fig. 8（b）It is these
Original image corresponding testing result on the pedestrian detection model of the present embodiment.From Fig. 8（a）And Fig. 8（b）As can be seen that the greatest extent
The problem of pipe blocks is not resolved, but the model of the present embodiment obtains very well on INRIA pedestrian detection data sets
Pedestrian detection result.
In short, by showing the scheme that the present embodiment proposes in the experiment of INRIA and Caltech pedestrian detection data sets
It is really effective, the accuracy and rapidity of pedestrian detection can be significantly improved.
In addition to the implementation, the present invention can also have other embodiment.It is all to use equivalent substitution or equivalent transformation shape
At technical solution, fall within the scope of protection required by the present invention.
Claims (10)
1. a kind of pedestrian detection method based on deep learning, which is characterized in that include the following steps：
Video image input extraction network to be detected is generated characteristic pattern by the first step；
Second step will extract the characteristic pattern input area proposal network that network generates, and region is recycled to propose that method detection most has
It may be the region of pedestrian, generate pedestrian candidate person and the corresponding score of pedestrian candidate person；
Third step determines whether pedestrian candidate person is real pedestrian using trained decision Tree algorithms.
2. a kind of pedestrian detection method based on deep learning according to claim 1, it is characterised in that:The extraction net
Network uses PVANet networks, the PVANet networks to have 14 layers, and wherein three first layers are convolutional layer, and centre is two groups of initial layers,
Every group of initial layers include the identical initial layers of four structures, and last three layers are full articulamentum；The output of the full articulamentum is area
Propose the input of network and decision tree classifier in domain.
3. a kind of pedestrian detection method based on deep learning according to claim 2, it is characterised in that:Single initial layers by
First, second, third branch forms, and first branch is made of one 1 × 1 convolutional layer, and second branch is by one 1
The convolutional layer composition of × 1 convolutional layer and one 3 × 3, the third branch is by one 1 × 1 convolutional layer and two 3 × 3
Convolutional layer forms.
4. a kind of pedestrian detection method based on deep learning according to claim 3, which is characterized in that single initial layers
Generating characteristic pattern, the specific method is as follows：Last layer generate characteristic pattern by respectively be passed to initial layers three branches, then by
The characteristic pattern of these three branches output is transferred into an articulamentum, finally enters next layer, becomes next layer of input feature vector figure.
5. a kind of pedestrian detection method based on deep learning according to claim 4, it is characterised in that：It is carried in the region
It discusses in network, for the input feature vector figure that PVANet networks generate, a sliding window is used and is generated on each width characteristic pattern
Multiple features connected entirely, the feature each connected entirely include Liang Ge branches, and one of branch is scs layers, another branch
It is cds layers；Described cds layers is used for generating pedestrian candidate person, including the coordinate of pedestrian candidate person's central point and the pedestrian candidate
The width and height of person；Described scs layers is used for generating the corresponding score of pedestrian candidate person；The pedestrian candidate person generated by cds layers
Score corresponding with the pedestrian candidate person generated by scs layers is transported to decision tree classifier and is trained and detects.
6. a kind of pedestrian detection method based on deep learning according to claim 5, it is characterised in that：Single sliding window
The region motion of different scale and different aspect ratios can be predicted simultaneously.
7. a kind of pedestrian detection method based on deep learning according to claim 6, it is characterised in that：The sliding window
When the region motion of four scales of prediction and four length-width ratios, 4*4 region motion will produce.
8. a kind of pedestrian detection method based on deep learning according to claim 7, it is characterised in that：The sliding window
4*4*4 output is generated at cds layers, and 2*4*4 output is generated at scs layers.
9. a kind of pedestrian detection method based on deep learning according to claim 8, it is characterised in that：Using
RealBoost algorithms train the method for decision tree as follows：
1. given training set,
（x1, y1）...（xi, yi）...（xN, yN）
Wherein, yiIt is feature vector, and i=1 ..., N；
2. in the starting stage, Weak Classifier is numbered and number is denoted as j, according to（1）Formula determines each Weak Classifier
Weight,
（1）
Wherein, WjFor the weight of Weak Classifier, H is the number of Weak Classifier；
3. carrying out n times to Weak Classifier to train to obtain training data, training is numbered and number is denoted as n, then basis
（2）Formula obtains a probability Estimation,
（2）
Wherein, Pn（y）For the probabilistic estimated value of Weak Classifier, N is the frequency of training of Weak Classifier；
4. basis（3）Formula calculates the true Distribution value of Weak Classifier,
（3）
Wherein, fn（y）For the actual value of Weak Classifier, R is set of real numbers；
5. in the training process, according to（4）Formula obtains the weight of Weak Classifier,
（4）
6. after each iteration, the weight of all Weak Classifiers is normalized again so that total weight of all Weak Classifiers
Equal to 1, strong classifier is finally obtained, according to（5）Formula obtains strong classifier,
（5）
Wherein, N is the frequency of training of Weak Classifier.
10. a kind of pedestrian detection method based on deep learning according to claim 9, which is characterized in that determine real row
The method of people is as follows：Classify to the pedestrian candidate person of generation first with trained decision Tree algorithms, then according to row
People's candidate classification preset threshold value, when the probability that the pedestrian candidate person in characteristic pattern is pedestrian is less than preset threshold value
When, then pedestrian candidate person classification is pedestrian, and the otherwise classification is not pedestrian.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810082310.0A CN108460336A (en) | 2018-01-29 | 2018-01-29 | A kind of pedestrian detection method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810082310.0A CN108460336A (en) | 2018-01-29 | 2018-01-29 | A kind of pedestrian detection method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108460336A true CN108460336A (en) | 2018-08-28 |
Family
ID=63239171
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810082310.0A Pending CN108460336A (en) | 2018-01-29 | 2018-01-29 | A kind of pedestrian detection method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108460336A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109409309A (en) * | 2018-11-05 | 2019-03-01 | 电子科技大学 | A kind of intelligent alarm system and method based on human testing |
| CN110263712A (en) * | 2019-06-20 | 2019-09-20 | 江南大学 | A kind of coarse-fine pedestrian detection method based on region candidate |
| CN110837769A (en) * | 2019-08-13 | 2020-02-25 | 广州三木智能科技有限公司 | Embedded far infrared pedestrian detection method based on image processing and deep learning |
| WO2020191668A1 (en) * | 2019-03-27 | 2020-10-01 | Hangzhou Fabu Technology Co. Ltd | Proposal processing method and related products |
| CN111797769A (en) * | 2020-07-06 | 2020-10-20 | 东北大学 | Small target sensitive vehicle detection system |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102103690A (en) * | 2011-03-09 | 2011-06-22 | 南京邮电大学 | Method for automatically portioning hair area |
| CN102496001A (en) * | 2011-11-15 | 2012-06-13 | 无锡港湾网络科技有限公司 | Method of video monitor object automatic detection and system thereof |
| CN105243356A (en) * | 2015-09-10 | 2016-01-13 | 北京大学 | Method of building pedestrian detection model and device and pedestrian detection method |
| CN105975929A (en) * | 2016-05-04 | 2016-09-28 | 北京大学深圳研究生院 | Fast pedestrian detection method based on aggregated channel features |
| CN106897738A (en) * | 2017-01-22 | 2017-06-27 | 华南理工大学 | A kind of pedestrian detection method based on semi-supervised learning |
| CN107240087A (en) * | 2016-11-01 | 2017-10-10 | 北京深鉴智能科技有限公司 | Object detection system and method |
| US20170358086A1 (en) * | 2015-09-30 | 2017-12-14 | Fotonation Limited | Method and system for tracking an object |
-
2018
- 2018-01-29 CN CN201810082310.0A patent/CN108460336A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102103690A (en) * | 2011-03-09 | 2011-06-22 | 南京邮电大学 | Method for automatically portioning hair area |
| CN102496001A (en) * | 2011-11-15 | 2012-06-13 | 无锡港湾网络科技有限公司 | Method of video monitor object automatic detection and system thereof |
| CN105243356A (en) * | 2015-09-10 | 2016-01-13 | 北京大学 | Method of building pedestrian detection model and device and pedestrian detection method |
| US20170358086A1 (en) * | 2015-09-30 | 2017-12-14 | Fotonation Limited | Method and system for tracking an object |
| CN105975929A (en) * | 2016-05-04 | 2016-09-28 | 北京大学深圳研究生院 | Fast pedestrian detection method based on aggregated channel features |
| CN107240087A (en) * | 2016-11-01 | 2017-10-10 | 北京深鉴智能科技有限公司 | Object detection system and method |
| CN106897738A (en) * | 2017-01-22 | 2017-06-27 | 华南理工大学 | A kind of pedestrian detection method based on semi-supervised learning |
Non-Patent Citations (6)
| Title |
|---|
| CORNELIU ZAHARIA 等: "PVANet optimization for person detection", 《2017 INTERNATIONAL CONFERENCE ON OPTIMIZATION OF ELECTRICAL AND ELECTRONIC EQUIPMENT》 * |
| JEROME FRIEDMAN 等: "Additive logistic regression: a statistical view of boosting", 《ANNALS OF STATISTICS》 * |
| SANGHOON HONG 等: "PVANet: Lightweight Deep Neural Networks for Real-time Object Detection", 《网络在线公开: HTTPS://ARXIV.ORG/ABS/1611.08588》 * |
| WEICHENG SUN 等: "Deep learning based pedestrian detection", 《2018 CHINESE CONTROL AND DECISION CONFERENCE》 * |
| 孙炜程: "基于深度学习的上下文相关的目标跟踪", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 徐科 等: "《金属表面质量在线检测技术》", 30 October 2016, 冶金工业出版社 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109409309A (en) * | 2018-11-05 | 2019-03-01 | 电子科技大学 | A kind of intelligent alarm system and method based on human testing |
| WO2020191668A1 (en) * | 2019-03-27 | 2020-10-01 | Hangzhou Fabu Technology Co. Ltd | Proposal processing method and related products |
| US10909417B2 (en) | 2019-03-27 | 2021-02-02 | Hangzhou Fabu Technology Co.. Ltd. | Proposal processing method and related products |
| CN110263712A (en) * | 2019-06-20 | 2019-09-20 | 江南大学 | A kind of coarse-fine pedestrian detection method based on region candidate |
| CN110263712B (en) * | 2019-06-20 | 2021-02-23 | 江南大学 | Coarse and fine pedestrian detection method based on region candidates |
| CN110837769A (en) * | 2019-08-13 | 2020-02-25 | 广州三木智能科技有限公司 | Embedded far infrared pedestrian detection method based on image processing and deep learning |
| CN110837769B (en) * | 2019-08-13 | 2023-08-29 | 中山市三卓智能科技有限公司 | Image processing and deep learning embedded far infrared pedestrian detection method |
| CN111797769A (en) * | 2020-07-06 | 2020-10-20 | 东北大学 | Small target sensitive vehicle detection system |
| CN111797769B (en) * | 2020-07-06 | 2023-06-30 | 东北大学 | Small-target-sensitive vehicle detection system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP6830707B1 (en) | Person re-identification method that combines random batch mask and multi-scale expression learning | |
| CN110599448B (en) | Migratory learning lung lesion tissue detection system based on MaskScoring R-CNN network | |
| CN109919108B (en) | Remote sensing image rapid target detection method based on deep hash auxiliary network | |
| CN108460336A (en) | A kind of pedestrian detection method based on deep learning | |
| CN109344736B (en) | Static image crowd counting method based on joint learning | |
| WO2020102988A1 (en) | Feature fusion and dense connection based infrared plane target detection method | |
| CN106778595B (en) | Method for detecting abnormal behaviors in crowd based on Gaussian mixture model | |
| CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
| CN103366602B (en) | Method of determining parking lot occupancy from digital camera images | |
| CN106940816A (en) | Connect the CT image Lung neoplasm detecting systems of convolutional neural networks entirely based on 3D | |
| CN108830188A (en) | Vehicle checking method based on deep learning | |
| CN110033473B (en) | Moving target tracking method based on template matching and depth classification network | |
| CN104537647B (en) | A kind of object detection method and device | |
| CN112287788A (en) | Pedestrian detection method based on improved YOLOv3 and improved NMS | |
| CN111160249A (en) | Multi-class target detection method of optical remote sensing image based on cross-scale feature fusion | |
| CN108492272A (en) | Cardiovascular vulnerable plaque recognition methods based on attention model and multitask neural network and system | |
| CN109241982A (en) | Object detection method based on depth layer convolutional neural networks | |
| CN106980858A (en) | The language text detection of a kind of language text detection with alignment system and the application system and localization method | |
| CN110826379B (en) | Target detection method based on feature multiplexing and YOLOv3 | |
| CN108875624A (en) | Method for detecting human face based on the multiple dimensioned dense Connection Neural Network of cascade | |
| CN107507170A (en) | A kind of airfield runway crack detection method based on multi-scale image information fusion | |
| CN110716792B (en) | Target detector and construction method and application thereof | |
| CN110543906B (en) | Automatic skin recognition method based on Mask R-CNN model | |
| CN110008899B (en) | Method for extracting and classifying candidate targets of visible light remote sensing image | |
| CN110349167A (en) | A kind of image instance dividing method and device |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180828 |
|
| RJ01 | Rejection of invention patent application after publication |