Research On Vision-based Object Detection And Scene Segmentation For Traffic Scene Understanding

Vision-based object detection and scene segmentation play an important role in traffic scene understanding.This dissertation is mainly grounded upon two projects: the National Natural Science Foundation grand project “Key techniques and integrated validation platform for autonomous land vehicle” and the National Natural Science Foundation key project “Research on the task related visual attention mechanism and the non-safe driving state analysis method”.This dissertation focuses on vision-based object detection and scene segmentation for traffic scene understanding.We propose several effective methods for autonomous driving systems and advanced driver assistant systems(ADAS).Main contributions and innovations of this dissertation are listed as follows:(1)We propose a single-layer convolution based proposal generation approach.The approach can effectively avoid exhaustive object searching across the given image.Besides,the approach is able to reduce the number of samples that need to be classified.In this way,the subsequent object detector can achieve higher recall rate and lower false positive rate.(2)We propose a multi-population genetic algorithm based approach for adaptive network structure adjustment.For a given base network,the approach can adaptively determine the number of kernels and feature channels for each layer in a heuristic way.It can avoid empirical handcraft network construction to some extent and make a trade-off between the performance and the network size.Besides,we propose a TLD(TrackingLearning-Detection)based multi-frame fusion strategy,which can effectively build connections between objects in adjacent frames,and make use of the temporal information to improve the recall rate and reduce the false positive rate of the object detection model.(3)We propose a generalized Haar filter based network compression method.The method makes use of the simple structure and strong representation of the generalized Haar filter to normalize the weights of convolutional neural networks.It can not only reduce the consumption of storage resource and computing resource,but also improve the generality of the model.Besides,we propose a local regression strategy that uses the“divide and conquer” framework to divide the global regression task into several easier local regression tasks.The strategy can improve the performance of the small object detection and reduce the requirement of the network size for object detector.(4)We propose a gradient fusion based multi-task training framework,which makes use of the training error of each channel to balance the back propagation of the gradient.The framework can not only improve the convergence rate of multi-task training,but also perform fine-tuning for each channel when the training error is small.Based on this framework,we propose a dimension stretching based method for scene segmentation,which takes full advantages of the dimension information to perform up-sampling for feature maps and improves the performance of scene segmentation.Furthermore,we propose a cascaded convolutional neural network based object detection method.The method can reject most negative samples layer by layer,and achieve better object detection performance in the end.We perform several experiments on recently published datasets and autonomous vehicle platforms,the experimental results validate the effectiveness and rationality of our methods.