Optimization And Implementation Of Vehicle Detection Algorithm In Complex Environment

With the continuous increase in car ownership, traffic congestion and other problems have become increasingly serious.Numerous studies have been done on Intelligent Transportation System, which aims to make full use of road resources to ease the traffic pressure. Statistics of traffic flow is a core job to evaluate traffic congestion. By traffic congestion evaluation, road resources can be reasonablely allocated. In complex environment,inaccurate vehicle detection makes a direct influence on traffic flow statistics,which leads to affection to the performance of Intelligent Transportation System.Thus, study on vehicle detection in complex environment is of great value theoretically and practically.Complex environment refers to environment of rain, snow, night time and vehicle occlusion. High detection rate has been achieved in rain,snow and night time. This thesis focuses on environment of vehicle occlusion.First,a thorough analysis is made to current occlusion detection algorithms,which can not guarantee real-time,detection rate and multiple vehicle occlusion handling simultaneously.Then,an occlusion detection algorithm based on rectangular template is proposed.This method first screening out obvious occlusion by area ratio.To area ratio lower than threshold,feature point is used for further detection to improve detection rate. When occlusion is detected,vehicles are devided by peripheral feature points which enables the method ability of handling multiple vehicle occlusion. Simplicity of rectangle template makes detection and segmentation so fast that can meet the real-time requirement. In addition, occlusion detection method relys on the integrity of the foreground object contour.To the problem of low recall rate of ViBe foreground detection algorithm, an improved algorithm is proposed by combining gray model with color space model and using the pixel level and frame-level background model update strategy. This dramatically improves the results of vehicle detection.Experimental results show that proposed algorithm can achieve detection rate of 91.02% and is capable of handling multiple vehicle occlusion. The real-time part test on a Core 2GHz processor with an average of 320 × 240 resolution video monitor, result shows each frame process in 23ms. Compared with Heidari fast algorithm.error detection rate is reduced by 50% and capable of handling multiple vehicles occlusion, while each frame processing time with an increase of only 2ms. Compared with the high detection rate CHIU algorithm, detection rate decreased only by3.03%, but each frame processing time is reduced by 50%.The proposed algorithm can meet the real-time requirement with high detection rate and can handle multiple vehicles occlusion which can be applied to the actual traffic flow statistics and other scenes.