Research On Vision-based Pedestrian And Road Sign Detection

Pedestrian detection and road sign detection are important technologies used in autonomous driving and assisted driving.They are also the hot research areas in the field of computer vision and pattern recognition.As an important part of the generic target detection task,a large number of scholars have proposed many methods and conducted targeted research on the difficulties involved.Deep learning performs well in detection tasks in recent years,while the improvement of hardware computing power has made such methods real-time performance.However,due to the characteristics of the target itself and the presence of a large number of interfering factors in natural scene,it’s still difficult to achieve fast and accurate detection.Occlusion and small scale are the main reasons for the reduced detection accuracy.Meanwhile,many models are unable to meet the real-time requirements due to the long computation time.This thesis focuses on these problems and tries to meet the real-time performance while improving the detection accuracy.The main research of this thesis is as follows:A pedestrian detection method combines head and overall information is proposed to address the current problem of reduced detection accuracy caused by occlusion and too small scale in pedestrian detection.This method first constructs a feature pyramid with a multi-layer structure.Then,by fusing the feature maps output from different substructures of this feature pyramid,it provides targeted feature information for head detection and pedestrian detection.Finally,the non-maximum suppression algorithm is improved to combine the pedestrian head information with the overall information to obtain final detection results.The effectiveness and generalization performance of proposed method are verified by experiments on public datasets.A road sign detection method combined with semantic segmentation is proposed for the occlusion problem in this task.This method starts from two aspects: the correlation of feature channels and the sensitivity of the model to the spatial location of the target.First,the backbone network is improved by using squeeze-and-excitation module,so that it can effectively distinguish the importance of each feature channel and has feature channel adaptive calibration capability,which makes the network pay more attention to the effective part of the target.Then,this thesis combines semantic segmentation and object detection to form a multi-task joint learning,so as to improve the accuracy of model prediction location and reduce the interference of background information.Finally,the outputs are filtered to obtain the road sign information in images.Experiments on publicly datasets show the proposed method improves the detection accuracy of the model to a certain extent and increases the computational complexity only slightly.