Research On Vehicle Flow And People Flow Detection Based On Kinect

With the rapid development of social economy and artificial intelligence,traffic flow detection and people flow detection have become a hot issue in the field of video detection,and the market demand is strong.People,vehicles,and roads are the three key elements in the analysis of intelligent transportation systems.The key to solving the problem of traffic jam is to fundamentally solve the contradiction between people,vehicles and roads.In the world coordinate system,the road is relatively stationary,but the vehicle and the people are moving relative to the ground,which results in the traffic flow and the people flow.At present,traditional methods use color cameras to detect traffic flow and people flow,which are easily affected by light,shadows,and so on.Especially in nighttime environment,there is insufficient illumination,low visibility,and less color image features to achieve accurate detection.Those traditional methods cannot detect targets without light.In recent years,with the appearance of Kinect,the acquisition cost of the depth image has been greatly reduced,and using Kinect can obtain the depth data of the lost color image,which can make up for the deficiency of using the color image.Therefore,we change the method of data collection and use Kinect depth image to solve the problems of traffic flow detection in nighttime environments and people flow detection.Firstly,introduce the structure and working principle of Kinect,and analyze the Kinect development tools: Kinect for windows SDK and OpenNI,then achieve depth image acquisition.And then describe the causes of the depth image noise and hole,analyze the relationship between image stabilization and the number of holes,and study the denoising and inpainting of depth image.In addition,in order to overcome the problems of low accuracy,poor stability,and incapability of valid vehicle identification based on color camera-based vehicle detection methods in nighttime environments,we propose a nighttime traffic detection algorithm based on deep virtual loop.We preprocess the depth image to generate a Motion Depth Map(MDM)and a Hole Depth Map(HDM).In addition,we set up virtual loop on MDM and HDM,and use integral images to generate corresponding one-dimensional motion signals,and weight them to realize the expression of vehicle motion characteristics.The one-dimensional signal obtained by combining the strength signal and the hollow signal is used to realize the traffic flow count.The vehicle target position is obtained by synthesizing the count signal range to detect whether the vehicle is changing lanes or crossing lanes,Thus,the error of vehicle statistics can be reduced,the geometric features of vehicle targets are calculated by the algorithm,and the vehicle model is effectively identified by SVM.Experimental results demonstrate that the proposed method can detectand count vehicles accurately in nighttime environments with 99.75% and 99.25%recognition rates for the one-lane and two-lane scenarios respectively.It can also accurately classify types of vehicles in terms of light and heavy vehicles with 99.80%accuracy rate.The average time of processing one frame is only 7 ms.Then,in order to avoid the complexity of target detection and tracking,a deep virtual loop algorithm traffic detection algorithm was proposed.By setting a virtual loop to extract feature signals in the depth image,we analyze the variation of the count signal when a pedestrian passes a virtual loop and normalize the count signal.Count signal denoising is achieved by analyzing the noise law of the counting signal and its cause.Then we obtain the people motion direction and achieve the people flow count,and analyze the system's various parameters and application scenarios.Experimental results demonstrate that the correct rate of the proposed method in single channel can reach 97.20%.Finally,for the problem that the Traffic flow count of the color camera is vulnerable to the influence of light and shadow,which results in poor detection effect,a traffic flow detection algorithm based on depth map and target tracking is proposed.By analyzing the methods of foreground detection and background modeling,we propose a method of jointing depth interception and average background modeling to achieve foreground detection.In the process of head detection,the local extremum region is obtained by random point and overflowing filling method.The head region detection is realized by extracting the head-shoulder characteristic descriptor with lower dimension as well as using threshold limit and SVM classifier.By using target characteristics to match and track,we use multiple counting lines to obtain the flow of people,and propose a method to obtain the information of pedestrian height.Experimental results demonstrate that the proposed method can detect and count pedestrian accurately in different fields of view with99.30% recognition rates,and the average relative error of height measurement is less than2.59%,the average time of processing a single image only take 18.31 ms.