Crowd Object Detection Based On Deep Learning

With the acceleration of the urbanization process,public safety issues have become increasingly prominent.As the main security control method,video surveillance plays an important role in protecting the safety of people's lives and property.And as the country vigorously carries out urban construction projects such as "safe cities" and "smart cities",the security field urgently needs to realize intelligent monitoring with the help of deep learning target detection technology on the basis of manual monitoring.The target detection technology based on deep learning theory has achieved great success in recent years,but the detection effect of small dense targets such as crowds needs to be improved.This paper proposes some improvement schemes for the difficulties related to crowd target detection.The main research work of this paper includes:(1)According to the difficulties of the large number of crowd targets,small targets and high density,adopt the head target detection method,and locate the individuals in the crowd by detecting the head targets of the crowd.The SCUT-HRAD data set is selected as the experimental data set in this paper,and the single-stage detection algorithm YOLOv3 with faster detection speed is selected as the main algorithm in this paper.On the SCUT-HRAD data set,the detection accuracy of the YOLOv3 algorithm for crowd targets is 85.26%.The detection speed FPS of the three GTX1080 Ti environments is 94.3,and the training time of the neural network model is 6 hours and 34 minutes.(2)Aiming at the problem that the ratio of the a priori frame and the real head target do not match,a method to modify the Anchor ratio of the a priori frame at multiple scales is proposed to achieve the effect of accelerating network convergence.Training the neural network model on 3 GTX1080 TiGPUs,the training duration was shortened from 6 hours and 34 minutes to 6 hours and 21 minutes,and the training duration was shortened by 3.3%.(3)For the problem of low detection accuracy of dense crowds,a crowd target detection algorithm based on attention mechanism is proposed,and the crowd density map generated by the crowd density estimation algorithm CSRNet is successfully integrated into the feature map in the form of attention mechanism,so as to improve the crowd The effect of target detection accuracy.The experimental results show that the detection accuracy rate mAP on the SCUT-HEAD data set is increased from 85.26% to 86.35%.At the same time,due to the addition of the CSRNet network structure,the detection speed of the algorithm is reduced from 94.3FPS to 92.1FPS in the case of 3 GTX1080 TiGPU.The reduced speed still meets the requirements of real-time detection,so this article believes that it is valuable to sacrifice a small part of the speed in exchange for the improvement of the accuracy of crowd target detection.On the premise of satisfying real-time detection,in the field of security monitoring,a GPU with a lower configuration can be selected according to the actual situation to complete the crowd target detection task.