Research On Algorithm Of Vehicle Detection In Static Images

This paper summarized the algorithms of vehicle detection in static images used in recent years, and analyzed their characteristics. And then a new vehicle recognition algorithm using local features is proposed. This algorithm contains two main steps: hypothesis generation(HG) step and hypothesis verification(HV) step.In the hypothesis generation step, a color transform model based on Karhunen-Loeve (KL) transform is proposed. Utilizing the model, All pixels in recognition images can be transformed from RGB color space to a 2-D color space (s,t). On this 2-D color space, vehicle pixels and non-vehicle pixels will respectively concentrate on detachable areas. hypothesis generation using the model can effectively reduce searching range in hypothesis verification step and decrease running time of vehicle detection algorithm.In the hypothesis verification step, a subspace vehicle recognition algorithm using local features is proposed. The contents includes pretreatment of vehicle samples, local feature selection, feature extraction by principal component analysis (PCA) and design of the nearest neighbor classifier. For decrease contrast difference of vehicle samples produced by different scene and different light condition, the proposed algorithm pretreat vehicle samples by histogram equalization firstly, then three subregion are selected as local feature by analyze the influence on each subregion of vehicle samples of lighting conditions and background noise. Afterwards, three eigenspaces are generated by using PCA to reduce the dimension and extract the feature. Finally the same feature extraction operation to recognition image is done, the nearest neighbor classifier for vehicle classification.Testing results demonstrate that by using the proposed algorithm of hypothesis generation and hypothesis verification, the vehicle detection can be realized with a strong robusticity and high accuracy.