Research On Visual Saliency Detection And Semantic Analysis In Driving

Visual saliency detection while driving is an important part in the Intelligent Vehicle System(IVS) and vehicle active safety system．The purposes of this research is giving real-time alarm to thedriver from driver assistant system, and reducing traffic accidents to protect personal property damage,improving traffic safety and transport efficiency, and alleviate issues such as energy consumption andenvironmental pollution. And now, this field has been attracting worldwide researchers’ attentionincreasingly.The traffic environment visual perception is the core of this paper, through analysis of humanvisual information process, the traffic sign detection and recognition, lane detection are discussed.Main content in this research has been summarized as follows:(1) Inspired by neural signal processing of biological visual system, the study of the static anddynamic salient target detection method are proposed. The first algorithm is based on toplogicalindependent component analysis and golabal color contrast, the salient region of multi-scale andmulti-channel in single image is detected, the experimental results show this algorithm is valid innatural and artificial image; the second method is for the moving target detection, through the analogof visual local relative entropy, the local Kullback Leibler divergence is used, and a model similar toMarkov is build to detect moving targets in dynamic scence, the compare with other methods provesthe effectiveness of our method.(2) A method for high-level traffic scene semantics understanding is proposed. Firstly, an imageenhancement method based on the atmospheric reflection model is proposed, which is effective forfog and haze scene, sencondly, a feature description idea from Bag of Words (BOG) model is made,and the Probability Latent Semantic Analysis (PLSA) method is used to build the mapping functionbetween image and senmatic, the test accuracy in datasets and real scene is more than80%.(3) A joint model based on high-level’s semantics and low-level’s probability is proposed. Withthe different semantic scenes, the attention will be guided to different objects. Depending on the hightask, the features of underlying data makes basis to select the correct object for attention. Firstly, amodel for retina is built, through a Gabor wavelet, the filter’s character of retina is simulated, and then,an attention model with bottom and upper task is built with probability cooperation, thirdly, a specifictask of searching traffic sign is test, the result demonstrates the model is valid.(4) A novel salient objects detection model in traffic environment detection method is presented.This model contains two parts, the traffic sign and road line. For traffic sign, firstly, opponent color and radial symmetry are used to describe traffic sign, and then the center-surround algorithm is usedto enhance the sign area, secondly, the Log Gabor and Phase Congruency is used to get signs’ contour,and thirdly, the SVM is used to recognize the signs’ meaning, some experiments based on GermanTraffic Sign Recognition Benchmark are made, the results show that the proposed approach hashigher detection rate than traditional approach. For road line, firstly, the Gabor wavelet is used to getthe image’s edge, secondly, the fast Hough transform is employed to collect the edge with samedegree, thirdly, with suitable threshold, the short segements and flat curve are filtered. The test is on anormal video, which shows the proposed algorithm can satisfy the safe and real-time driving.(5) An intelligent vehicle driving assistant device based on visual sensor is developed. Thehardware is designed including video collecting, algorithm proceeding and display. The software isbased on OpenCV platform (Open source Computer Vision) by Intel. Which provides both lanedepartures reminds and traffic sign recognition function. An on-board experiment has beenimplemented using this system, and the result verifies the effect of the approaches in this paper.In summary, this paper focus on saliency extracting, semantic analysis, the traffic imageenhancement and active attention, and the road lane and traffic sign detection algorithm based onthese research are proposed, in the end, a real test of traffic saliency detection and semantic analysisare made in driving, the results contribute practical and useful significance to both theory reseach andengineering application in related fields.