Research On Pedestrian Detection And Tracking Based On Convolutional Neural Network

With the rapid development of artificial intelligence,the automotive industry is undergoing tremendous changes.Driverless cars become the development trend of automotive industry.The environment perception system is responsible for detecting the environment around the driverless car,identifying obstacle and pedestrian,which plays an important role in ensuring the safe driving of the driverless cars.Pedestrian detection and tracking algorithms as one of the core technologies of environmental perception system become a research topic that has received much attention in recent years.Based on convolutional neural network,this paper studies the pedestrian detection and tracking algorithm and proposes a pedestrian detection and tracking method with good real-time performance.First,considering the particularity and real-time requirements of pedestrian detection,this paper analyzes the object detection model that based convolutional neural network and proposes a new pedestrian detection method by improving YOLOv3-tiny network.The algorithm improves the network's ability to extract features by adding convolution layers,and enhances the network's ability to detect small-sized pedestrian by fusing shallow and deep features of the network model.In order to further improve the detection accuracy,this paper uses the loss function that fuses the Gaussian model to predict the uncertainty of the bounding box to improve the positioning ability.The experimental results show that the improved network has better accuracy and real-time performance.Then,considering the timing information of the video,this paper proposes a pedestrian detection model with convolutional LSTM.Compared with pictures,the information in the video is highly correlated in the time dimension,and this correlation is helpful to detection.Convolution LSTM is a special kind of recurrent neural network,which can obtain video context information while extracting image features.Therefore,each frame of the video is associated during detection by embedding the convolutional LSTM in the convolutional neural network,and the context information is used to detect video.The experimental results show that the model embedded convolution LSTM can effectively reduce the missed rate and improve the detection accuracy.Finally,this paper implements a real-time pedestrian tracking algorithm based on the pedestrian detection algorithm,combined with the Kalman filter algorithm and the Hungarian algorithm.First,the Kalman filter algorithm is used to track the target quickly,then the pedestrian detection algorithm is used to detect the input image.After the pedestrian detection result is matched with the tracking result through the Hungarian algorithm,the Kalman filter algorithm is used to modify the matched result and get the final tracking result.The experimental results show that the pedestrian tracking algorithm in this paper can achieve fast and accurate tracking.