Implementation Of An Embedded Vehicle Counting Method Based On Deep Learning

Recently,with the increasing number of vehicles,there raises traffic issues dramatically in some traffic trunk road of cities.To ease the traffic issues mentioned above,an efficient and intelligent traffic management system is required to manage the traffic situation.One of the areas in traffic management is to monitor the traffic at the intersection of roads.The current monitoring methods require human to count the number of vehicles manually in each direction of intersection during rush hour.However,the human resource is expensive and the staffs will get tired and drop in their efficiency during a long time monitoring,which leads to losing vehicles as numerous vehicles with various colors in the intersections.To overcome these shortcomings,several related work has been presented,and one of them is vehicle counting on the road.Note that the accuracy of the conventional technique drops dramatically with the vehicle running on lane lines.In order to achieve the vehicle-flow counting,we propose a vehicle counting method based on the Single Shot MultiBox Detector(SSD)that will track each blob within successive frames.The thesis is composed of three sections,which are organized as follows.(1)In the aspect of moving vehicle detection,the traditional detection algorithm such as background differential method and deep learning method respected by Regions with CNN features(R-CNN)are introduced respectively.In this paper,we propose a counting method based on SSD and train the model with KITTI database.The SSD has the capacity to extract meaningful features of blob of vehicles,and predict their bounding boxes with centroids and classification of vehicles from each frame.(2)In terms of vehicle tracking and counting,combined with the centroid and bounding box of the moving vehicle obtained by the SSD algorithm,the tracking and counting methods used in this paper are proposed.We propose a circle searching algorithms to predict the possible position of a moving vehicle in the next frame and seek the best point in the relevant region to reduce the target searching range.Euclidean distance is used to measure the distance between their centroids.SURF feature matching is adopted to solve the discontinuity problem of vehicle detection.Furthermore,given that the virtual coil method is easy to implement,we utilize this method to count the number of moving vehicles with fast speed.(3)Finally,the experiment results from six traffic scene videos including three videos with more complicated traffic scenes and three ones with little complicated traffic scenes are provided to illustrate that the effectiveness of the proposed method on counting vehicles with above 82%accuracy of real-time processing speed on a GTX1080Ti GPU and useful information for traffic surveillance.The article uses the NVIDIA Jetson TX2 development board as an embedded platform and transplants the algorithm into TX2 to lay the foundation for further development of the proposed algorithm.