The Multi-scale Detection Method For The Dense Crowd Target Detection

With the rapid development of transport,more and more public places been full of the crowd gathered are emerging,which suround business district,transportation hub.The implied security besetting the government and businessmen is a long-term problem,timely access to accurate crowd information is the key to solve this problem.From the field of computer vision,timely access to crowd information is to improve the target detection technology,to achieve high speed and high precision of the realistic requirements.From the market's deployment of current population density target detector,the low average accuracy mostly affect the large-scale application of the target detector.Therefore,improving the average precision of the target detector is the key to solve the security hidden danger in the crowded places.The main work of the thesis:(1)For the problem that the detection speed of DPM(Deformable Part Model)detector is too slow,this paper proposes a parallel optimization algorithm to accelerate the detector.The algorithm uses OpenMP(Open Multi-Processing)parallel technique to divide the input data of HOG(Histogrrams of Oriented Gradients)feature pyramid process and model matching process,and make full use of hardware calculation of multi-core CPU(Central Processing Unit)Resources,so that each core deal with the divided computing tasks,thereby improving the computational efficiency.The algorithm also uses the SIMD(Single Instruction Multiple Data)technique to invoke the AVX(Advanced Vector Extensions)instruction to speed up the 32-D dot product operation between the filter and HOG features.The optimized DPM detector is called Faster DPM detector.Experimental results show that the parallel algorithm has a high speedup,compared to DPM detector,Faster DPM detector has a higher detection efficiency.(2)Aiming at the problem of low accuracy of target detection technique in dense population environment,this paper proposes a head-target detector based on deep convolution neural network called AnyScale detector.The detector can not only preserve the real image information,but also improve the detection precision by using the convolution layer as the detection layer,which can overcome the traditional target detector's limitation of the size of the input image.According to this special application environment of dense crowd target detection,the detector uses the concept of “ anchor ” to parameterize the input tag information to make the target detector applicable to the dense target environment.The experimental results show that the AnyScale detector has higher average detection accuracy than the other main target detectors in dense population target detection environment.(3)Aiming at the problem that the detection accuracy of single target detector is low,a multi-scale detector based on ensemble learning theory is proposed in this paper.The detector is a heterogeneous ensemble learner and consists of two individual learners,the Faster DPM detector and the AnyScale detector.The combined module of multiscale detector maintains the advantages of the two individual detectors,eliminates the redundant prediction target,and outputs the most reasonable detection result.So that on the coarse scale,it depends on the Faster DPM detector;on the fine scale,it depends on the AnyScale detector.The experimental results show that the multi-scale detector has a significant improvement in the detection accuracy compared to a single detector.