Research On Crowd Counting And Density Estimation Based On Computer Vision

With the rapid growth of the population and diversifying into human activities,crowded scenes can be seen everywhere,such as airports,stations,stadiums and so on.High-density crowds always bring unexpected security risks.Extreme congestion,stampede and other violent incidents have brought a lot of challenges to traditional public safety management systems.As a result,crowd counting,as a typical crowd statistics task,has received a significant amount of attention in the past several years.Researchers have investigated such technique on real-world applications,e.g.,public security,pedestrian monitoring and disaster management.Automated vision-based solutions normally have a characteristic of low cost and high efficiency.Therefore,computer vision-based crowd analysis is becoming an increasingly important task.With the continuous improvement of science and technology,crowd counting has gradually completed the transition from simple crowd counting to spatial density estimation.In this article,we will discuss the above together.With the continuous development and technology innovation,compared with the traditional computer vision,the deep learning has achieved unprecedented success in performance and deployment.This paper deeply analyzes the technical pain points of actual deployment and then proposes targeted measures depending on deep learningbased solutions.The main work of the paper is as follows:(1)This paper compares the different types of regression models in the field of crowd counting and density estimation,and analyzes the improvement of the performance brought by value quantization trick.The above observation shows the fact that feature quantification can improve the model performance.To solve the problem on the destruction of local information and the curse of dimensionality,this paper proposes a model component based on semantic feature quantization,which can be able to independently deploy into the existing model and learn relative knowledge with an End-to-End training.Experiments demonstrate that the accuracy(Such as MAE and RMSE)of the method exceeds the baseline model and such quantification strategy shows superior performance.(2)This paper lists three typical model training strategies(Class-Balance,Hard Example Mining,and Curriculum Learning)in the field of machine learning,and points out that the series of strategies can improve model performance.With the ability of point supervision introduced by the Bayesian-based method for crowd counting tasks,we take a closer look at the results.This paper proposes an improved bayesian loss which could guide the model to learn on the different types of samples.The experiment proves that such method can effectively guide the model to take an efficient training course.(3)With the quality of results and quality of service,this paper discusses the process from research to production on crowd counting and density estimation.Firstly,this paper summarizes the researches on lightweight models in related computer vision fields and proposes a lightweight CNN-based network for crowd counting and density estimation tasks.Secondly,the paper summarizes the famous model-independent lightweight solutions,such as parameter quantification and network pruning.And then the usability of such solutions based on TensorFlow deep learning framework were discussed.Finally,this paper goes deeper in the quality of service and proposes a quantifiable method for the crash risk of the crowd which visually describes the risk in crowd scenario.