CN109635667A - A kind of vehicle detecting system based on Guided Faster-RCNN - Google Patents

Abstract

The invention discloses a kind of vehicle detecting systems based on Guided Faster-RCNN, comprising the following steps: obtains picture from monitor video, detect and save picture to the picture using Guided Faster-RCNN algorithm；Picture is uploaded onto the server；Designated mobile phone is sent by the picture of preservation, provides technological guidance for the real-time detection of vehicle.Designed vehicle detecting system is with good performance, and substantially increases detection speed using GPU acceleration, can satisfy the requirement of real-time.

Description

A kind of vehicle detecting system based on Guided Faster-RCNN

Technical field

The present invention relates to a kind of vehicle detecting systems based on Guided Faster-RCNN algorithm, belong to computer vision Field.

Background technique

With the raising of standard of mass living, the quantity of automobile gradually increases, and road traffic pressure also increasingly increases.At present Most of vehicle lacks detailed vehicle model information, and hides the monitoring of road traffic with this, brings to traffic management department Many troubles.Therefore, to solve such issues that, people start with various new technologies come such issues that solve, intelligent transportation system System also comes into being in this background.The purpose of target detection is that movement mesh is picked out from the background of differing complexity Mark, and separating background, to complete the follow-up works such as tracking, identification.One as target detection technique of vehicle testing techniques Particular studies field obtains always the attention of domestic and international researcher for many years, and constantly has new research achievement to go out It is existing.For vehicle detection as the main problem in video monitoring, studying related algorithm has very the performance for improving monitor video Important meaning.

The method of currently used video encoder server mainly has optical flow method, frame-to-frame differences method, Background difference and is based on machine The methods of study.Three kinds of detection methods of front are all the conventional vehicle checking methods based on image processing techniques, however this A little algorithms require vehicle that cannot have and block very much otherwise tracking accuracy decline greatly with very time-consuming and algorithm, or even will cause The disadvantages of target is lost.From 2012, deep learning causes the extensive concern of people, and takes in image recognition and detection Obtained good recognition effect.The present invention has studied application of the deep learning in target detection, and will be in target detection Guided Faster-RCNN algorithm applies in vehicle detection.One kind that the present invention designs is based on Guided Faster- The vehicle detecting system of RCNN, designs and realizes and apply to Guided Faster-RCNN algorithm in vehicle detection, structure Build and realize vehicle detecting system, and vehicle detection is accelerated to realize real-time detection by GPU so that even if In complicated scene, the real-time of this system is also met the requirements, and no matter all has very big advantage, tool in speed or precision Standby good performance.

Summary of the invention

In view of the above-mentioned problems, the purpose of this law, which is to provide one kind, is based on Guided Faster-RCNN vehicle detecting system, It solves in traditional detection method information to calculate quite time-consuming and block greatly very much since vehicle has and tracking accuracy is caused to decline The problem of even target is lost, proceeds from the reality, real-time detection vehicle, to solve the problems of existing road traffic monitoring Technological guidance is provided, has the characteristics that detection speed is fast, detection efficiency is high, feedback is convenient.

Based on the vehicle detecting system of Guided Faster-RCNN, structure include RPN network, Fast-RCNN network, Shared convolutional layer, difficult example Mining Strategy training, non-maxima suppression.

The RPN network is added to an adaptive pool layer after the convolutional layer of original RPN, and RoI is navigated to Corresponding region in Feature Map, and down-sampling is that fixed size feature is passed to full articulamentum again.Training process is passed using reversed Algorithm and stochastic gradient descent are broadcast, entire training process is the process of an End-to-end.Training RPN when, ignore it is all across The more Anchor of image boundary prevents exceptional value from bringing difficult amendment error term, causes training that cannot restrain.RPN most After later layer convolutional layer Conv5 output, carried out using random corrective unit (Randomized Leaky ReLU, RReLU) non- Linear activation.

The Fast-RCNN network recommends frame, one Fast-RCNN of stand-alone training using the target that RPN network generates Detect network.RReLU activation primitive, activation primitive output access multiple dimensioned area-of-interest pond layer are used in Fast-RCNN (Spatial Pyramid RoI Pooling Layer, SPP), to adapt to more dimensional variations and deformation.In addition in training When convolutional layer in using Batch Normalization and Selective Dropout technology accelerate convergence, to avoid excessively quasi- It closes.Original Fast-RCNN detection network uses Pooling layers of RoI, improved detection Web vector graphic one of a single layer Multiple dimensioned version (SPP), is adapted to more scales and deformation, and SPP layers of initialization is initialized using gaussian random.Inspection When survey, convolutional layer has been completed with SPP layers of shared calculating in RPN network, avoids duplicate calculating, and improve detection Energy.

The shared convolutional layer gets up RPN network and Fast-RCNN detection network associate, and keeps network structure not Become.RPN recommendation network is initialized using trained Fast-RCNN detection network, then re -training RPN.Re -training When fixed RPN whole convolutional layer be used for sharing feature, only finely tune latter half of Pooling layer exclusive of the RoI of RPN, classification layer and time Return layer network layer, obtains the candidate region the RPN recommendation network of training completion.Keep whole convolutional layers again in Fast-RCNN It is fixed with SPP layers, the full articulamentum of Fast-RCNN latter half is only finely tuned, the Fast-RCNN target network of training completion is obtained Network.Finally, the shared of whole convolutional layers is realized in two networks, can extract common feature, both ensure that detection speed That spends is efficient, in turn ensures learning process and detection process is the process of complete End-To-Endd.

The difficult example Mining Strategy training is by the way of augmenting difficult example negative sample, the difficulty of Guided Faster-RCNN Example mining algorithm is shown in algorithm 1.It is obtained first using original training set (including the random negative sample of positive sample and part) training first It is negative then to excavate all difficult examples on more sample sets (including whole positive samples and more negative samples) for beginning depth model Sample (erroneous detection sample) finally obtains optimal depth model to update initial model.

The difficult example mining algorithm of 1 Guided Faster-RCNN of algorithm:

Input: original training set R0 caches Tt, sample set R

Output: Y (R)

Initialization: Tt=R0

1) using caching Tt training depth model Yt=Y (Tt).

If 2) classify on bigger sample set R without difficult example negative sample (erroneous detection sample) using Yt, Yt is returned.

If 3) step 2 does not return, expand caching.Difficult example the negative sample A(Yt, R that Yt is sorted out on sample set R) addition Into caching Tt, new caching is Bt.

4) training caching Tt+1, obtains depth model Yt+1=Y (Tt+1).

5) 2 to 4 step n times are repeated, optimal depth model Y(R is obtained)=Y(Tt+N).

The non-maxima suppression is used to remove the candidate that some overlap ratios generated in RPN network are high, confidence level is low Frame.CUDA technology passes through the thread that can largely execute parallel and accelerates to algorithm, and by GPU dynamic dispatching and execution, solves It is simple using CPU to calculate quite time-consuming problem to carry out non-maxima suppression, greatly reduce the time required for operation.

The vehicle detecting system based on Guided Faster-RCNN is designed Guided Faster- and is realized RCNN algorithm is used in road monitoring vehicle detection, constructs and realize vehicle detecting system, and by GPU to vehicle Detection is accelerated to realize real-time detection.

The invention has the following advantages that the frame base of Guided Faster-RCNN in the present invention in Faster-RCNN On plinth, shallow-layer study guidance is added, maximum possible avoids the over-fitting occurred in study.It improves in network simultaneously Some Key Technologies, improve learning ability, be a complete End-to-end study and detection process.Guided Faster-RCNN uses the training method based on difficult example Mining Strategy, can further improve performance, and the overwhelming majority is calculated in GPU Upper completion, is greatly reduced detection time.

Detailed description of the invention

Fig. 1 is the sampling flow chart in RPN network training process；

Fig. 2 is the training flow chart of RPN network Yu Fast-RCNN network；

Fig. 3 is the schematic diagram of the low candidate frame of confidence level in non-maximum restrainable algorithms removal RPN network；

Fig. 4 is vehicle detecting system overall block flow diagram.

Specific embodiment

The present invention will be further explained below with reference to the attached drawings.

The present invention is achieved by the following technical solutions: sampling flow chart as shown in Figure 1, original RPN last Layer convolutional layer Conv5 (1) accesses an adaptive pool layer (2) afterwards, and it is right in Feature Map (4) that RoI layers (3) are navigated to Region is answered, and down-sampling is that fixed size feature is passed to full articulamentum (5) again, in each convolution mapping position, exports the position The target classification score of n candidate region of upper different scale and different length ratio and target position return boundary (6).The knot Center in structure on RoI layers (3) for each sliding window all carries out the anchor point of n different scale, different length-width ratios Sampling, for the Feature Map of x × y size, will there is x × y × n times sampling process.

Training flow chart is as shown in Fig. 2, one region RPN recommendation network of stand-alone training first, for obtaining a series of mesh It marks candidate region and accordingly detects score.An adaptive pool is added to after the last layer convolutional layer Conv5 of original RPN Layer, so that different size of image is mapped as the Feature Map of fixed size.It is defeated in the last layer convolutional layer Conv5 of RPN After out, samples random amendment linear unit and carry out nonlinear activation.Recommend frame followed by the target that the RPN of the first step is generated, The Fast-RCNN of one multiple dimensioned characteristic of stand-alone training detects network.RReLU activation primitive is used in Fast-RCNN, activates letter Number output access multiple dimensioned area-of-interest pond layer, to adapt to more dimensional variations and deformation.Third step passes through shared volume Lamination gets up two network associates, keeps the network structure of two networks constant.It is trained with second step when finely tuning RPN Fast-RCNN detects network to initialize the region RPN recommendation network, and re -training RPN, fixed RPN whole convolutional layer is for altogether Feature is enjoyed, the latter half of exclusive RoI layer of RPN, classification layer are only finely tuned and returns layer, finally obtains the RPN candidate regions of training completion Domain recommendation network.When finely tuning Fast-RCNN, whole convolutional layers and SPP layers of fixation are kept, only finely tune Fast-RCNN latter half Full articulamentum.Two networks are alternately finely tuned, and all training process are finally completed, and realize that whole convolutional layers are total in two networks It enjoys, can extract common feature.

Non-maxima suppression schematic diagram is as shown in figure 3, all recommendation frames (left side) that RPN network generates use non-maximum The recommendation frame (right side) selected after restrainable algorithms.The candidate frame number that RPN network generates is more and overlap proportion is high, more in order to eliminate Remaining frame removes the time that some overlap proportion are high, confidence level is low using non-maxima suppression algorithm according to the height of confidence level Select frame.

System general flow chart as shown in figure 4, from monitor video obtain demand picture, utilize Guided Faster-RCNN Algorithm detects the picture.Guided Faster-RCNN algorithm is returned under if obtained result is no vehicle One picture is detected；The picture is uploaded onto the server if obtained result is to have vehicle.Server main thread is used In display video, server sub thread is for sending a picture to designated mobile phone.When a thread, corresponding user is saved in figure After piece, it will send a piece of news push to corresponding mobile phone app, point can be seen after opening PUSH message on webpage The picture uploaded onto the server provides excellent technique guidance for the real-time detection of vehicle.

Claims (3)

1. a kind of vehicle detecting system based on Guided Faster-RCNN, including it is RPN network, Fast-RCNN network, shared Convolutional layer, difficult example Mining Strategy training module, non-maxima suppression module, it is characterised in that: the RPN network is original An adaptive pool layer is added to after the convolutional layer of RPN, training process uses back-propagation algorithm and stochastic gradient descent, whole A training process is the process of an End-to-end, after the last layer convolutional layer Conv5 output of RPN, using with machine maintenance Positivity unit carries out nonlinear activation；Fast-RCNN the network recommends frame, stand-alone training one using the target that RPN is generated A Fast-RCNN detects network, and RReLU activation primitive is used in Fast-RCNN, and activation primitive output access is multiple dimensioned interested Pool area layer；The shared convolutional layer gets up RPN network and Fast-RCNN detection network associate, and keeps network knot Structure is constant, alternately finely tunes RPN network and Fast-RCNN network, detects network using trained Fast-RCNN to initialize RPN recommendation network, then re -training RPN, realizes the shared of whole convolutional layers in two networks；The difficult example excavates plan Slightly training module is by the way of augmenting difficult example negative sample, the initial depth model obtained first using original training set training, All difficult example negative samples are excavated on more sample sets then to update initial model, finally obtains optimal depth mould Type；The non-maxima suppression module is used to remove the candidate that some overlap ratios generated in RPN network are high, confidence level is low Frame passes through the thread that can largely execute parallel using CUDA technology and accelerates to algorithm, and by GPU dynamic dispatching and execution.

2. the vehicle detecting system according to claim 1 based on Guided Faster-RCNN, which is characterized in that described RPN network be in the stand-alone training stage, using trained Fast-RCNN detect network come initialize the region RPN recommend net Network, then re -training RPN and fixed RPN whole convolutional layer are used for sharing feature, only finely tune the latter half of exclusive RoI of RPN Pooling layers, classification layer and recurrence layer network layer, finally obtain the candidate region the RPN recommendation network of training completion.

3. the vehicle detecting system according to claim 1 based on Guided Faster-RCNN, which is characterized in that described Fast-RCNN network in keep convolutional layer and SPP layer fixation, the only full articulamentum of fine tuning Fast-RCNN latter half, finally Obtain the Fast-RCNN target network of training completion.