Unmanned (3) Pedestrian Tracking Algorithm

Student ID: 2002 12 10873

College: School of Electronic Engineering

This paper mainly introduces the pedestrian tracking algorithm needed in unmanned driving.

Computer vision Kalman filter particle filter mean shift of unmanned environment perception embedded in bovine nose

What are the pedestrian tracking algorithms used in unmanned driving?

Mosaic ox script

Pedestrian tracking has always been a difficult problem in the visual field, and the research of pedestrian tracking algorithm is influenced by external factors such as complex practical application environment, occlusion and pedestrian posture change. Pedestrian tracking algorithm models are mainly divided into generation model and discrimination model.

(A) mode of production

Generation model is a pedestrian tracking model established by learning the characteristics of pedestrian targets online, and then using this model to search the target area with the smallest error, thus completing the tracking of pedestrians. The algorithm only considers the characteristics of pedestrians and ignores the background information when constructing the model, and fails to effectively use all the information in the image. Among them, the classical algorithms mainly include Kalman filter, particle filter, mean shift and so on.

(1) Kalman filtering algorithm

Kalman filtering algorithm is an optimal linear recursive filtering algorithm which calculates the minimum mean square error based on the pedestrian's state equation and observation equation, and predicts the pedestrian's trajectory change by recursive pedestrian's motion state.

First, set the initial parameters and read the video sequence. Then, background estimation is performed to generate an initialized background image. Then the video sequence is read in turn, and according to the estimated background of the previous frame and the data of the current frame, the foreground target of the current frame is obtained by using Kahnan filtering algorithm. Then, the connectivity of the foreground object is calculated to detect the trajectory of the moving object. The classical Kalman filtering algorithm can only track pedestrians moving in a straight line. Later, scholars improved the Kalman filter algorithm, which can track pedestrians with nonlinear movement, with less calculation and real-time tracking, but the tracking effect is not ideal.

(2) Particle filtering

The core of particle filter is Bayesian inference and importance sampling. Particle filter can be used in nonlinear non-Gaussian model, because Bayesian inference adopts Monte Carlo method, and its probability is expressed by the frequency of events at a certain time point. The posterior probability distribution of the whole model is approximately represented by a group of particles, and the state of the whole nonlinear non-Gaussian system is estimated by this representation. Importance is given different weights through the confidence of particles, particles with high confidence are given greater weights, and similarity is expressed by the distribution of weights.

(3) Mean shift

Mean shift algorithm belongs to kernel density estimation method. There is no need to know the prior probability, and the density function value is calculated by the characteristic space of sampling points. The target model is described by calculating the probability of pixel eigenvalue in the target area of the current frame, and the candidate areas are described uniformly. A similarity function is used to express the similarity between the target model and the candidate template, and then the candidate model with the largest similarity function value is selected to get the mean shift vector about the target model, which represents the vector of the target moving from the current position to the next position. By iteratively calculating the mean drift vector, the pedestrian tracking algorithm will eventually converge to the actual position of pedestrians, thus realizing pedestrian tracking.

Discriminant model

Discriminating model is different from generating model, and pedestrian tracking is regarded as a binary classification problem. The pedestrian and background information in the image are extracted and used to train the classifier. The pedestrian is separated from the image background by classification, and the current position of the pedestrian is obtained. Taking the pedestrian region as a positive sample and the background region as a negative sample, the positive and negative samples are trained by machine learning algorithm, and the trained classifier is used to find the region with the highest similarity in the next frame to complete the pedestrian trajectory update. Different from the generation model, the discriminant model uses not only pedestrian information but also background information, so the tracking effect of the discriminant model is generally better than that of the generation model.

(1) Tracking Algorithm Based on Correlation Filtering

? Kernel correlation filtering (KCF) algorithm is a classical tracking algorithm based on correlation filtering, which has good tracking effect and tracking speed. This is because it uses cyclic shift to generate samples, trains the classifier with the generated samples, calculates the similarity probability map of pedestrians in the current frame and all candidate targets in the next frame through Gaussian kernel function, finds the candidate target with the largest similarity probability map, and obtains the new position of pedestrians. In order to improve the tracking accuracy, KCF algorithm uses HOG features to describe pedestrians, and combines discrete Fourier transform to reduce the computational complexity.

(2) Tracking algorithm based on deep learning.

In recent years, deep learning has made great achievements in image and speech, so many researchers combine deep learning with pedestrian tracking and achieve better performance than traditional tracking algorithms. DLT is a pedestrian tracking algorithm based on deep learning. Using automatic depth model encoder, the pedestrian model is obtained from a large-scale pedestrian data set through off-line training, and then the pedestrian is tracked online to fine-tune the model. Firstly, the candidate pedestrian target is obtained by particle filtering, and then the predicted position of the pedestrian is obtained by automatic encoder, that is, the candidate pedestrian target position with the largest output value. The MDNet algorithm proposed by 20 15 adopts the method of domain training. For each category, a separate fully connected layer is used for classification, and all layers before the fully connected layer are shared for feature extraction. The HCFT algorithm proposed by 20 17 uses deep learning to train a large number of calibration data, and obtains a powerful feature expression model. Combined with the tracking algorithm based on correlation filtering, it is used to solve the problems of few pedestrian samples and insufficient network training in online tracking. In addition, features are extracted by deep learning, and tracking algorithms are realized by data association methods, among which the most famous methods are JPDAF and MHT.