Traffic Video_Based Realtime Vehicle Driving Status Recognition Algorithm Research

Realtime vehicle driving status recognition research is an important task for intelligentizing traffic video monitoring. By means of vehicle driving status realtime recognition, it can make the computer explain the detail of vehicle driving status to the traffic supervisors automatically, and what will happen also can be predicted according to the vehicle current driving status, then alarm are sent to the abnormal vehicle under the specific traffic rules, in the end the safety traffic can be achieved.This paper put emphasis on researching the algorithms of recognizing realtime vehicle driving status, and the main contributions are as follows:Firstly, this paper produces an algorithm for achieving vehicle driving status recognition based on Hidden Markov Models(HMMs).The algorithm firstly takes some preprocessings to the trajectory sequences including abandon the not enough length trajectory sequences, linear smooth filtering and Least Square fitting, that to guarantee the usability of acquired trajectory sequences; Secondly, motion trajectory direction angle features from trajectory sequences has been extracted and on this basis a direction angle region partition algorithm is produced to generate the observation sequences, which will determine the different trajectory patterns that acquired by vehicle realtime various driving status according to the change of direction angle regions; Finally, by the multiple observations based Baum-Welch algorithm, the optimal HMM model parameters of each trajectory pattern in specific traffic scene can be trained, then through acquiring trajectory segment sequences and matching with above trained HMM models, the recognition of realtime vehicle driving status is achieved. Experiment results are listed to demonstrate the effectiveness and stability of this algorithm.Secondly, with the Baum-Welch algorithm hill-climbing searching problem may bring the local early convergence results, and in order to improve the recognition accuracy of vehicle driving status, this paper lucubrates the genetic algorithm, and the influence of the genetic parameters to the global convergence is pointed out. Then to solve this matter a new method called genetic population diversity based adaptive algorithm is produced. This algorithm adapts the genetic parameters according to the chromosome genetype and phenotype which are defined to determine the population diversity. From experiments the HMMs trajectory model parameters trained by our adaptive genetic algorithm perform better than the traditional Baum-Welch algorithm in the optimal solution,average convergence speed and recognition accuracy. Thirdly, in order to optimize the process of HMMs parameters training, and improve the average convergence speed, this paper produced a hybrid genetic algorithm and Baum-Welch algorithm to train the HMMs parameters which are based on the wide scope searching ability of the proposed adaptive genetic algorithm and the local optimum searching ability of the Baum-Welch algorithm.This hybrid algorithm is applied to train the HMM trajectory models in experiment and results show that it can perform better than the other algorithms in the optimal solution, average convergence speed, recognition accuracy, and is finally more qualified to meet the realtime vehicle driving status recognition.