Research On Pedestrian Detection With Low Parameter Based On Deep Learning

Pedestrians are vulnerable groups in road traffic.For intelligent vehicle,it is of great significance to quickly and accurately detect pedestrians on the road.After years of development,the pedestrian detection algorithm based on deep learning has been able to detect pedestrians on the road well,but there are still many problems that need to be solved,such as poor detection capacity of long-distance pedestrians and occluded pedestrians,slow algorithm speed,and too many parameters.This paper further studies the pedestrian detection algorithm based on the low parameter object detection algorithm YOLOv3-tiny.To analyze the problems of high miss rate of YOLOv3-tiny and the poor detection capacity of long-distance pedestrians and occluded pedestrians,an improved algorithm YOLOv3-tiny-D is proposed: in the extraction of prior box,the original Kmean algorithm is replaced with Kmeans ++ algorithm;in the backbone network part,replace the original convolutional network structure with depthwise separable convolution to deepen the network depth.In order to evaluate the performance of the algorithm before and after the improvement,we tested it on the Caltech pedestrian dataset.The test results showed that the improved algorithm's miss rate in various scenarios were lower than the original algorithm,and the overall missed detection rate was 60.97%,a decrease of 4.3%,and a decrease of 8.64% when pedestrians are heavily occluded.There is still a problem with the improved algorithm YOLOv3-tiny-D,that is,poor ability to detect long-distance pedestrians.In order to enhance the algorithm's ability to detect long-distance pedestrians,a new algorithm YOLOv3-tiny-DM is proposed: In the detection part,an multi-level feature pyramid network based on depthwise separable convolution was proposed.The network consisted of eight feature pyramids with the same structure.The feature pyramid was also composed of depthwise separable convolutions.The feature pyramids were connected in series.The same feature pyramid was responsible for extracting features of different scales.Different feature pyramids were responsible for extracting features of different depths.Finally,to obtain feature pyramids with different depth features,the features of the same size obtained by different pyramids were merged.Then the fused feature pyramid was used for detection.The test results on Caltech Pedestrian dataset show that,the overall miss rate of this method is 59.38%,which is 30.98% lower than HOG method,and 3.21%?1.57% lower than deep learning method SA Fast-RCNN and MS-CNN,respectively.When detecting pedestrians with heavy occlusion and pedestrian with long distance,the miss rate of this method is 20.04% and 12.66% lower than SA Fast-RCNN method,17.27% and 8.25% lower than MS-CNN method.The detection speed is 34 ms / frame,meeting real-time requirements.