The Research On The Methods Of Detection Of Vehicles In Video Sequence

The monitoring of the dynamic changing scenes is an active frontier research areain recent years. The detection and recognition of moving targets in complex trafficscenes is the key technology in intelligent video surveillance. The result directly affectsthe subsequent trajectory tracking, behavior understanding and the processing of themonitored events.For the issue of detecting moving objects in real-world traffic scenes, we proposeda method using an improved algorithm of α-blending (the adaptive update algorithm ofbackground model). Compared with the Optical Flow algorithm and the GaussianMixture Model algorithm, the improved α-blending algorithm can better overcomeharsh conditions such as background disturbance, dramatic changes of the outsideilluminance and so on. It can also extract images of moving targets in the complexbackground effectively with higher efficiency.To detect the moving vehicles in the proposed method, the characteristic of theHistogram of Oriented Gradient (HOG) is used. First the image is divided into units andthe histograms of multi-scale orientations in each image unit are calculated. Then, thehistograms are normalized on each image region composed of four adjacent units andintegrated to form the feature set. Finally, the HOG feature set is used for the training ofa support vector machine (SVM) classifier for the classification of moving objects onthe road.On the basis of the theory and methods, we have implemented the system softwaredesign based on VC++. Experimental results show that with the improved α-blendingalgorithm and the SVM classifier trained using the HOG feature set, moving vehiclescan be detected and classified in real-world complex scenes. The effectiveness of thedetection algorithm and the software has also been validated through the experiments onactual traffic image sequences.