Research On Vehicle Detection Algorithms Based On Convolutional Neural Networks

Vehicle detection is the foundation of the autopilot system,it exerts crucial influence on the performance of the autopilot system.Compared with active sensor such as ladar,camera,which is vision sensor generates richer sets of information at a fraction of the cost.Therefore,there's a widespread research on vision-based vehicle detection.However,the existing vision-based vehicle detection can't be applied in production grade autopilot system,due to the unstable performance under different driving situations and environments.Thus there lies great significance in studying vision-based vehicle detection algorithm with robustness.The major work are as follows:(1)With the analysis of the existing algorithms for vehicle detection,convolutional neural network is used to achieve robust vehicle detection considering algorithm characteristics and research target.(2)Aiming at the weakness in small target detection of existing objective detection algorithms based on convolutional neural network,vehicle detection algorithms based on the fusion of second-order convolutional features is proposed in light of Faster RCNN.This method fuses multi-layer convolutional feature to detect vehicles and introduces second order response transformation into convolutional neural network for better performance.(3)Region proposal networks based on time context is proposed in view of time domain consistency to accelerate detection speed.Due to the detection results of previous frame and horizontal edges,this method generates regions of higher quality while fewer numbersto reduce calculation time and speed up vehicle detection.(4)A vehicle data set which contains different road situation,weather situation and driving situation is built.Parameter adjustment and experiments are performed on the data set.The experimental results indicate that,compared to Faster RCNN and existing detection algorithms based on fusion of multi-layer convolutional feature,the algorithm proposed in this paper shows fast detection speed and better performance in small target detection.Experiment on different situations shows this algorithm has better robustness than algorithms based on prior knowledge and classic machine learning.