Research Of Pedestrian And Car Detection Algorithm Based On Deep Neural Network

Pedestrian and car detection has been widely used in many fields,such as automatic drive,augmented reality,intelligent security,intelligent transportation,and it draws more and more attention from academies and industries.In recent years,many effective methods have been proposed by researchers home and abroad to solve this problem.But owing to the uncontrollable factors such as illumination change,scale change and partial occlusion in real life scenarios,car and pedestrian detection remains a challenging issue,its precision still needs to be further improved.What's more,the existing methods are with high computational complexity in general,and the detection speed also can't meet the needs of the practical applications.In this paper,we studied the pedestrian and car detection and carried out a systematical research,which optimize the existing models and promote the detection precision and speed.The major contribution of this dissertation is summarized below:(1)In this paper,we propose an efficient hierarchical fully convolutional network(EFCN)for real time object detection.EFCN employs the light weight MobileNet as the base network to significantly reduce the computation cost.Meanwhile,it detects objects in feature maps with multiple scales,and deploys a refining module on the top of each of these feature maps to alleviate the accuracy loss brought by the simple base-network.The refining module combines slow-resolution semantically strong features with high-resolution semantically weak features by deconvolution and skip connections,which constructs the expression feature that is more robust to car and pedestrian.We evaluate EFCN on the challenging KITTI dataset and compare it with the state-of-the-art methods in the view of accuracy,speed and model size.The results show that EFCN achieves a good balance between speed and accuracy,it has 25 × fewer parameters and is up to 31 × faster than Faster-RCNN while maintaining similar or better accuracy.(2)In order to further improve the detection precision and keep the speed of the EFCN,we add objectness detection module to the origin model.The objectness loss is used as one of the objective functions,which can introduce foreground and background information to the feature layer.What's more,the objectness score acts as a prior to guide the search of the objects,which can remove a large number of background samples in advance and significantly reduce the searching space of the objects.Through objectness detection,we address the foreground-background class imbalance of the single-stage detectors to some extent.Experimental results show that the optimized algorithm improves the detection accuracy and without the increase in the time of the calculation of the network.