Study On Pedestrian Detection Ahead Of Vehicle Based On Machine Vision

Pedestrians are the main participants in the road traffic.In order to decrease the collision accidents of vehicles and pedestrians incidence rate to the maximum extent,pedestrian ahead of vehicle collision warning system has become one of the most active research topics in the field of intelligent vehicles,the system can detect pedestrians in front of vehicle in time and it can make dangerous early warning to the possible pedestrian collision accidents the first time,thus pedestrian traffic accidents can be effectively avoided.According to the actual problems of pedestrian detection ahead of vehicle,combined with related principle of machine vision,the pedestrian detection method based on machine vision is studied and accurate pedestrian detection effect is realized.It provides certain technical support for the design of vehicle driving assistance system,such as pedestrian ahead of vehicle collision warning system.The main work of this thesis is as follows:(1)A semi-automatic method for fast building pedestrian ahead of vehicles dataset is proposed.Based on the feature of image color histogram,a feature probability model for semi-automatic labeling of pedestrian targets is established.Furthermore,the Bhattacharyya similarity function is used to realize the iterative matching of the pedestrian candidate regions.It improves the semi-automatic degree of tagging method and saves the time consumption of manual annotation.Finally,based on the software development platform of Tkinter and Python,a semi-automatic method of pedestrian annotation GUI interface is designed and implemented,and more than 30000 of the pedestrian image database is built through using the GUI interface.(2)A pedestrian detection method based on improved HOG-LBP feature and support vector machine is proposed.According to the LBP feature may not be suitable for all kinds of texture features,the improved LBP algorithm is proposed based on the traditional LBP feature extraction algorithm,and the improved HOG-LBP feature extraction method is also proposed.Finally,pedestrian candidate regions are generated by using the sliding window method,the improved HOG-LBP features of candidate regions are fed to the training and test of support vector machine classifier.The experimental results show that the proposed method has higher recall rate and lower miss rate than the original pedestrian detection method.(3)A pedestrian detection algorithm based on cascade multi feature extraction model is proposed.According to the number of candidate window for sliding window method is too large and bounding box prediction is not precise,the thesis adopts the method that generate pedestrian candidate regions by using convolutional neural network.The improved Faster RCNN for pedestrian detection is proposed based on Faster RCNN object detection method.Finally,after merging the improved HOG-LBP and the improved Faster RCNN proposed in this paper,a pedestrian detection method based on cascade multi feature extraction model is proposed.The complementary advantages of two methods are realized by using the classifier decision level information fusion and further improve the accuracy of pedestrian detection.Experimental results show that compared with using only one feature extraction method,the pedestrian detection method based on cascade multi feature extraction model has higher recall rate and lower miss rate,and the partial occlusion problem of pedestrian detection effect is also better.(4)A pedestrian detection real vehicle experiment verification platform based on embedded Linux system is designed.Pedestrian detection ahead of vehicle verification platform based on NVIDIA Jetson TK1 development board is built.In the verification platform,through the image in front of vehicle acquisition,image detection under the embedded processing platform and back-end output results,the real vehicle experiment verification of the pedestrian detection method based on multi feature model cascading is realized.