Traversability Analysis Method For The Front Of Vehicle Based On Monocular Vision

The vehicle population increases year by year, which leads to the continuously increasing risk of road traffic safety. As a result, the road traffic safety becomes an urgent problem to be solved by all circles. To reduce the risk of road traffic safety, vehicle safety driving assistance technology has been widespread concern. As an important research content of the safety assistant driving, the traversability analysis for the front of vehicle plays increasingly significant role in every aspect and reduces the risk of traffic.Traversability analysis for the front of vehicle has been doing a further study at home and abroad, including technical breakthroughs, as well as research bottlenecks, which is aimed at reducing the road traffic safety risk. With the deep learning as a guide, a traversability analysis method for the front of vehicle based on convolutional neural network is presented in this thesis, which can provide a theoretical basis for intelligent vehicle navigation and driver assistance safety, simultaneously reducing traffic accidents effectively.Aimed at the structured traffic environment, a method for traversability analysis for the front of vehicle based on monocular vision sensor is presented in this thesis, on the basis of analysis and summary of traversability analysis for the front of vehicle at home and abroad. Firstly, based on hierarchical design of the embedded system, an image acquisition system on-board is presented, to provide rich samples for the traversability analysis for the front of vehicle. The OMAP3530microprocessor is adopted in the hardware platform to achieve the image acquisition function. Secondly, for the sake of reducing the effect of illumination on samples, an improved Gamma correction algorithm is introduced, to achieve different levels of illumination compensation and enhance image quality. Furthermore, subjective and objective analyses are conducted to verify the results of illumination compensation. Based on a typical LeNet-5network, an improved convolutional neural network (CNN) is presented. Meanwhile, the Levenberg-Marquardt algorithm is used to adjust the weights of the CNN, avoiding the disadvantages of traditional BP algorithm. To validate the validity of the improved CNN, contrast experiments are carried out.Finally, the improved CNN is applied to the traversability analysis for the front of vehicle on expressway. By using sample images of the front of vehicle to train the network, a stable network structure is obtained and used for testing non-training samples. The testing results demonstrate that the traversability analysis method for the front of vehicle presented in this paper can effectively achieve object identification in front of the vehicle, which can provide a theoretical basis for the safety vehicle driver assistance and intelligent navigation.