Research On Moving And Multi-scaled Object Detection And Tracking

Intelligent transportation systems(ITS)use,besides other means,optical image sensors and the state-of-the-art in computer vision.These systems have an important significance for enhancing traffic efficiency and reducing traffic accidents.Special ITSexamples are advanced driver assistance systems(ADAS)or intelligent traffic surveillance systems(ITSS).Object detection and tracking plays a key role for ITS,here often as a technology for detection,localisation,and identification of traffic-relevant moving objects.This thesis focuses on vehicles.Being in motion and appearing at multiple scales(i.e.at different distances)are two main characteristics of vehicle targets in traffic scenes.Recently,much research is under way on this kind of targets in the context of ITS developments.Aiming at a robust detection and continuous tracking for moving and multi-scaled objects,this thesis thoroughly investigates vehicles in dynamic on-road scenes,challenging object appearance,variable image scales,as well as varying vehicle poses.Specifically,the main research contents addresses detection by adaptive ensemble learning,the use of a probabilistic graphical model in the spatial domain,mixed filter tracking,and tracking based on a deep graphical model.Major innovations in this thesis are summarized as follows:1.A scene-recognition micro-convolutional neural network(SR-MCNN)is proposed for the accurate recognition of vehicles in local scenes,including parameter adaptation for vehicle detection.2-layer scene recognition results are introduced into a vehicle detection framework that can reduce the false alarm rate for complex road scenes;for example,this increases the detection rate for dim(e.g.twilight)scenes.This adaptive detection framework improves the detection performance for moving and multi-scaled vehicles in dynamic and complex on-road scenes.2.A probabilistic graphical model is proposed in the spatial domain,called labelmoveable extended chain(LMEC).It ensures precise vehicle detection using license-plate elements.The proposed model can be used for effective capture of license-plate elements and accurate structure elucidation,even for license targets with a node missing,node deviation,or degradation by noise.It is verified that this method can be used for precise vehicle detection in various traffic scenes.3.A distance-relevant mixed-tracking(DRMT)algorithm is proposed that eliminates the nonlinearity caused by the rapid change of vehicle image scales.This also supports precise vehicle localization and continuous filter tracking.Experiments verify that this algorithm can be used for continuous vehicle tracking at critical ultra-close ranges.4.A novel graphical model is proposed in deep convolutional neural networks,called deep convolutional flow(DCF).By introducing the proposed model,we solve the problem of vehicle-pose identification and of estimating horizontal motion,so as to achieve robust vehicle tracking at long and middle range.All the above work has been verified by simulations and tests on extensive sets of real data.Experiments show that proposed methods can achieve a robust detection and continuous tracking in dynamic and complicated on-road scenes for moving and multiscale objects.