Pedestrian Detection Algorithm Design For Vehicle Assisted Driving

Pedestrian detection is a hot research topic in computer vision.In recent years,with the development of convolutional neural networks,various detection algorithms based on convolutional neural networks have emerged one after another.With more and more vehicles on the road today,the pedestrian safety issues that accompany them are becoming more and more prominent.High-performance pedestrian detection algorithms for vehicle-assisted driving have important practical significance,because it can effectively alert drivers are there pedestrians around,reduce the probability of accidents.The proposed pedestrian detection algorithm based on convolutional neural network is verified on the hardware system,the system can open up ideas for low-cost pedestrian detection system solutions for vehicle-oriented driving.In this thesis,the principle of object detection algorithm based on convolutional neural network is focused on,its redundant classification structure is removed,and obtains a neural network dedicated to pedestrian detection.Due to the large number of parameters of the network,the overall computational complexity is high,and the network must be streamlined to be suitable for use in vehicle assisted driving scenarios.The network is simplified in two aspects in this thesis:on the one hand,Binarizing the convolution kernel parameters of the network and input of the convolutional layer can effectively reduce the memory occupied by the network and accelerate the calculation of convolutions;on the other hand,the internal structure of the network can be tailored and replaced appropriately to effectively reduce network parameters and overall calculations.The two simplified networks are trained and tested respectively,and the better network is selected to impleament on hardware sysytem,make the fixed-point simulation by MATLAB to determine the required accuracy of each parameter in the hardware implementation.Finally,the video input/output module and data transmission module are designed and configured on the embedded hardware system,the convolutional acceleration module is designed and simulated to verify the pedestrian detection algorithm.The measurement accuracy of the pedestrian detection algorithm is based on the fact that the false positives per image is 0.1,the lower miss rate detection speed,the better performance of the algorithm.The algorithm designed in this paper is tested on the INRIA pedestrian data set.The miss rate is as low as 19%when the false positives per image is 0.1.With the support of a graphics processing unit,the detection speed can reach 66.7 frames/s.Compared with the referenced pedestrian detection algorithm,the overall performance of detection accuracy and speed is better.The pedestrian detection hardware system implemented on the embedded development board verifies the feasibility of implementing the algorithm in the embedded system.