Study On Algorithms Of Moving Vehicle Detection And Tracking Based On Monocular Camera

With the rapid development of pattern recognition,machine vision and artificial intelligence technology,the detection and tracking of front moving vehicle based on monocular camera has gradually become the hot-spot.It is the key technology and basis of the auxiliary driving system,even automatic driving.Moving vehicle detection and tracking is a process of detecting the motion in image sequences using color,boundary feature and color texture of the moving vehicle,and to extract the information such as the shape of the moving vehicle,and further to obtain the position,size and speed of the moving vehicle,distance and other movement state information,then to achieve the real-time tracking.However,it is very difficult to detect and track moving vehicles with accurate and high robustness due to the influence of shadowing,weather,intensity of illumination,complex background and uncertainty of target motion.The detection and tracking algorithm of forward moving vehicle based on monocular camera has been studied and discussed in this thesis.Some new algorithms are proposed to overcome the bottlenecks of the traditional vehicle detection and tracking technology.The main work and research results are as follows:(1)Because the traditional detection algorithm does not have the ability to detect the vehicle at night and the real-time performance of vehicle detection and tracking is flawed,this paper uses machine learning to extract the characteristics of five types of vehicles according to the characteristics of different convolution kernel.Then,a novel detection algorithm based on convolutional neural network is proposed.In order to extract the characteristics of the corresponding model,the convolution kernels corresponding to different features are firstly extracted by machine learning.Then,through training a large number of sample vehicles,the parameters of the convolution neural network are settled down,especially the BP neural network.Finally,according to the different characteristics,the same features with the same pixel-value tags are intelligently marked,and the information of target vehicle position is extracted.It can more quickly track the target vehicle.The simulation results show that the algorithms have good detection effect on forward moving vehicles and can effectively overcome the interference of unfavorable environmental factors,working out the bottleneck of traditional detection algorithms.The algorithm also has good robustness,and the detection results of this algorithm are closer to the effect of human visual classification.(2)For the real-time tracking of target vehicle is not ideal when tracking the target vehicle bythe novel algorithm based on the new convolution neural network.The moving vehicle tracking algorithm fusing super-pixels with target information block is proposed.Firstly,the target information blocks are extracted according to the existing detection results,and the target information blocks are divided into super-pixels.According to the concept of information entropy,the vehicle and background boundary are separated successfully,which effectively reduces the computation of the tracking algorithm.Then the block of target information based on the super-pixels is detected,and the target vehicle tracking effect is confirmed.The simulation results show that the tracking algorithm proposed in this paper has better real-time performance than the related algorithms,which can detect the target vehicle quickly accurately and in real-time.At the same time,the time cost of this algorithm is feasible for assistant driving system under high speed driving,and its detection error is very small,reaching the sub-meter level under 100Km/h high-speed.