Research On Some Problems Of Pedestrian Detection In Traffic And Their System Implementation

In recent years,with the rise of Driving Assistance System,pedestrian detection has also received more and more attention.The current pedestrian detection algorithms based on DNN have achieved good performance,but it is still far from actual demand in some cases.On the one hand,when a pedestrian is far away from the camera,he appears in the image with a smaller size and lower resolution,which will make the detector take him as the background and ignore him.On the other hand,pedestrians in traffic scenes are often blocked by vehicles,trees and other objects,which will greatly affect the recognition of pedestrians by the detector.In order to solve the above two problems,this thesis has conducted a deep research on small instances detection and occluded pedestrian detection in traffic scenes,and designed and implemented a pedestrian detection system.The specific work and main contributions are as follows:1.In the view of the problem that the pedestrian with small size has low resolution and is easy to be ignored in high-level convolution feature,a pedestrian detection method based on multi-scale feature mining is proposed.The region proposal model of this method uses convolution features of different levels to find the region of pedestrians.The experimental results show that the proposed method can do a better job in finding the pedestrian,reducing the number of pedestrians ignored in the region proposal stage,and improving the recall rate of pedestrian detection.2.In order to further solve the issue that the features of small instances are missing in the high-level convolutional features,we propose a detection method based on image segmentation technology,which predicts the location of pedestrian by predicting the center of pedestrian.The structure of proposed model is based on Unet network,and it enriches the feature information of the small target in the final feature map by fusing the low-level convolution feature in the high-level convolution feature.The experimental results indicate that our method gains the best detection results in both miss rate and precision.3.In the view of the problem that the occluded pedestrian in the traffic is diffcult to be detected,a pedestrian detection method based on attention guided mechanism is proposed.The proposed method takes the advantage of deep learning for sentiment recognition and utilizes the attention mechanism to guide the recognition network to foucs on the visible pedestrian's body features,while ignoring the features of occluded objects.At the same time,we also explore the effect of different attention mechanisms on occluded pedestrian detection.The experimental results demonstrate that the proposed attention guided neural network model has stronger robustness for the recognition of pedestrians with different occlusion degrees,and achieves the best performance on the subset of occluded pedestrians in multiple data sets.4.Based on the Tensorflow framework,pedestrian detection algorithm is implemented and successfully integrated into the server program.The client program based on VIM2 development board is designed and completed.The whole pedestrian detection system can run effectively and detect pedestrians accurately.