Research On Cars And Pedestrians Detection Model Based On Convolutional Neural Networks

The detection of road objects has widely applied to the field of intelligent transportation systems,advanced driver assistance system(ADAS)and auto-driving systems.For the application of advanced driver assistance system or auto-driving system,the object detection model requires high demand of speed and accuracy.Furthermore,the image capture device is located at a relative low position with narrow vision in advanced driver assistance system and auto-driving system,compared with the intelligent transportation system.Therefore more partly occluded objects will be captured,which are hard to be detected.Scale variation and illumination variation may also bring challenges.As computation capability of deep learning hardware upgrades year by year,the requirement of high speed or even real-time speed of the models is much easier to achieve.And deeper convolutional neural network models will be used at intelligent cars as long as higher detection precision is obtained.And for some areas don't rely on the speed of the models that much,like intelligent transportation systems,improving detection accuracy is more meaningful.A fully convolutional neural network based on YOLOv3 is used in this thesis.We implement multi-scale detection system based on the re-organized multi-scale feature maps.We train and test our cars and pedestrians detection model on the KITTI dataset,noticeable accuracy and high speed are achieved.We present a new algorithm called class-dependent k-means to deal with the imbalanced data problem.3-dimension information of the bounding box(width,height and class)is used while clustering,which eliminates the impact of the imbalanced data.The prior bounding boxes(or anchor boxes)generated by this algorithm will be more representative.And these priors will be more helpful for the imbalanced class(pedestrian in this case).The model starts off with better priors would be easier to learn the bounding box prediction task.After using our class-denpendent k-means clustering,better priors give an increase of up to 3.1% AP on pedestrian prediction.