Pedestrian Recognition And Optimization Based On Comparison Of Paired Features With Discrimination

With the improvement of information technology, the people’s requirements on the roadsand cars also raise, intelligent traffic came into being in this environment. In order to improvethe active safety of vehicles, driving assistance systems installing in vehicles to protect carsand pedestrians are widely used. This paper will propose an high performance detectalgorithm and its optimization for early warning for the pedestrain in driver assistancesystems.The main works of this article are as follow:In pedestrian detection feature extraction part, in order to overcome the shortcomingsthat histogram of gradient can not express pixel distribution in space and the over high ofdimensionality of the feature vector, Inspired from the phenomenon that animal visual systemuses color and gradient mode for simple comparison, we introduces a feature based on thecomparison of the gray value and gradient of the image block pairs. Different from previousworks which focus on a single information description of the object, we found acharacterization method that reflects the symbiotic relationship between the colour and thegradient orientation.Then this article will propose an improved algorithm and implementation of thisfeature,to solve the huge time consuming spend by global search in the feature pool. Thesynergistic effect of the two complementary modules increased the discriminative featuresearching speed in the vast feature space.Then use it for the enhanced object detectioncascade classifier. Experiments with a number of difficulties sample show that the detectionaccuracy is significantly improved and computational efficiency significantly improved.Finally, to get rid of the disadvantages of local optimization trend in heuristic searchalgorithm, we proposed a collaborative learning approach which is the combination of thesimulated annealing algorithm and heuristic incremental feature selection module,accelerating the speed of feature learning but also optimized the learning outcomes. Then usethe initial temperature fedback model in the cascade classifier, to adjust the constantlychanging of sample weight in the Adaboost algorithm, to stable annealing process,experiments show that the detection accuracy and the computational efficiency aresignificantly improved. The study has been made in scientific research and practices, good results can be widelyused in various applications of object recognition. And it also laid the theoretical foundationfor the field of machine vision of object recognition technology in the future.