Object Detection And Tracking For Intelligent Traffic

Intelligent Transportation System(ITS)integrates multidisciplinary theories and technologies such as computer,electronics,communications,automotive manufacturing and artificial intelligence to provide innovative modes of transport and intelligent traffic management services.The system can better improve urban traffic conditions,make it more coordinated,safe and humane,and improve travel efficiency.This paper deeply studies the problems of vehicles detection and tracking in vehicle monitoring and self-driving field based on the video data collected by visual perception system in intelligent traffic.We focus on the background extraction algorithms and establish the corresponding adaptive parameter adjustment mechanism.Feature fusion classification model and target tracking algorithm are proposed by using image feature.The calculation structure of deep learning parameters is optimized to control steering wheel angle.The main contributions of this paper are as follows:(1)A coupling models,"blinking model" and "target probability",are proposed to adaptively adjust the hyper-parameters of the visual background extraction model.In the intelligent traffic vehicle monitoring technology,the background model is firstly established according to the pixel level information to perform the object detection.Traditional background modeling sets fixed hyper-parameter based on experience,which cannot guarantee to be applicable to a variety of different scenarios.In order to solve this problem,a dual-model system is proposed to adaptively adjust the model parameters so that it can better adapt to a variety of scenes with background flicker,leaf shake,shadow interference and image noise.(2)A multi-feature fusion learning model is proposed and constructed to improve the object recognition rate and detection success rate of the vehicle-mounted camera.In self-driving,single image feature is difficult to meet the requirements of complex dynamic object detection.In order to solve this problem,we selected many cross-level features of the image to make decision,and added the gradient boosting model to create the feature fusion classifier to improve the object recognition rate and detection rate.(3)A new object tracking is proposed,which is based on the variable-scale adaptive searching window by using particle filter and a three-threshold template updating strategy.In the object tracking problem of intelligent traffic,due to the visual difference of the near and far objects,there are obvious changes in the scale and appearance of the target in the video stream for a long time.Therefore,this paper proposes a new image feature extraction method using multi-region fast color histogram integration,and uses the particle filter algorithm to find the best searching window size.We also propose a three-threshold target template updating strategy,which effectively resolves the unstable problems of the object tracking.(4)A parameter-compressed deep learning network structure is proposed to reduce the overall storage space of the model and improve the execution speed.Deep learning is the most commonly used method of image recognition,detection and other tasks.However,in the self-driving task,the parameters need to be reduced and the model size reduced to save the computing cost,so as to reduce the hardware requirement and improve the real-time performance guarantee.In response to this problem,according to the core idea of MobileNets network,the convolution neural network structure is further compressed and optimized to train the network which needs less storage space to execute the calculation.And this compressed neural network model is applied to control the steering wheel corner task.Under the premise of keeping the accuracy unchanged,the model calculation speed can be effectively improved.Finally,the paper summarizes the object detection and tracking in intelligent transportation studied in the full paper,and prospects the development direction and research methods in this field.