Pedestrian detection algorithms based on deep learning

Object detection has practical significance in many scenarios at present, and pedestrian detection is a necessary research field in object detection. Service robots, video surveillance, unmanned vehicles and drones all use pedestrian detection to perceive information of the surrounding environment, which can provide us an intelligent and safe society. Pedestrian detection has many problems and challenges, such as pedestrians usually in the complex environment, different pedestrian scales caused by various distances of the camera, the change of light, some objects are shaped like pedestrians in the background, pedestrians are obscured by each other in crowd and pedestrians are obscured by objects, all these difficulties are needed to solve. Many algorithms based on convolutional neural network(CNN) have emerged in object detection nowadays, which have achieved remarkable success and received widely attention. This paper focuses on the analysis and discussion of the feasibility and extensibility of the general framework Single Shot Multibox Detector(SSD) and You Only Look Once(YOLOv3) on pedestrian detection. However, after many experiments, I found that these frameworks have some shortcomings for pedestrian target, which lead to their poor performance in pedestrian detection. The regional candidate boxes of the original object detection algorithm have redundancy and deviation from pedestrian dataset, and the detection results of overlapping targets are a bit inaccurate. Therefore, this article redesigned the candidate boxes of the algorithm so that it can better fit the shape of pedestrian in the pedestrian dataset. On the one hand, I found that the logistic classifier used in the YOLOv3 framework usually can’t suppress difficult negative samples very well, like some objects in the background have similar appearance with pedestrians. Therefore, this article uses cascade boosting forest to distinguish pedestrians from difficult negative samples, which contributes to reduce the false positive results. On the other hand, for the SSD framework, the framework uses different level convolution feature maps to extract features, but the feature maps lack information about small scale pedestrians. Then I modify the network structure to improve the detection results, moreover, through optimizing loss function of the SSD algorithm, the detection results have big improvement in pedestrian occlusion. Finally, through the training and testing of the optimized networks on the Caltech pedestrian dataset, our models can better balance the detection results and speed in pedestrian detection