Research On Computer Vision Based Das Algorithm

The Driving Assistance System(DAS)is a system that continuously monitors the surrounding environment and sensing internal and external information of the vehicle during the driving,then feedbacks the danger signals to the driver and response the signals in time.An on-board system that takes over a vehicle.Object detection and recognition is an important branch of computer vision.With the development of deep learning and support of neural networks,real-time target detection can be achieved with extremely low missing and false rates of the detection.Different from autonomous driving technology,the driving assistance system has the participation of the driver.The driving assistance system can detect danger signals and remind the driver to avoid some traffic accidents caused by inattention or blind spot problems.The driving assistance algorithm purposing on the sensing of dangerous as the basis for making decision of DAS.First,this paper lists several target detection and recognition algorithms,and selects Faster R-CNN algorithm as the main algorithm used in this paper,which has both fast detection speed and high accuracy.The algorithm has a Region Proposal Network(RPN)for feature extraction by using shared convolutional layers of the convolutional neural network.The RPN network outputs the position information of Regions of Interest(RoI)and the result of performing a binary classification for each region.This approach maximizes the feature information extracted by the neural network,and also reduces the amount of computation of the classifier and the boundingbox model.Then,this paper uses the Deep Compression as the way of model compression.This method compressing model by sequentially prune,share weights,and quantize weights of the network model to obtain small-sized network model.In order to makes up for the lack of detection accuracy of small-sized network,this paper using Shortcut Connection combined with Complex Bypass structure in the compressed network model.To improve the detection accuracy of the compressed light-weight deep neural network,make the gratitude from output feature map of first layer flow to every layers after the first one increase the feature information received by the network.And using this network as the backbone network of Faster R-CNN algorithm to train and test the algorithm on the KITTI dataset.This reduced the amount of parameters of the compressed network significantly.Finally,this paper introduces the implementation methods and mathematical models of monocular and binocular vision ranging methods.And chooses the monocular vision ranging method as the ranging method used in this article which has simple mathematical model and low computational complexity.Then connect it to bounding-box model of the Faster R-CNN algorithm mentioned above.And input the coordinates of the midpoint on the two points at the lower end of the target positioning frame as the feature points input by the monocular visual ranging method to obtain a more accurate Ranging results.With the abilities of location and ranging of cars and peoples on the road,this algorithm can be used as the backbone algorithm of the driving assistance algorithm and used in the driving assistance system.