Project #1 Tracking pedestrian trajectories in outdoor videos

Automatic pedestrian tracking in real time videos is an active research field in computer vision, mainly due to its potential application in automotive safety systems among others. With increasing advances in machine learning and intelligent autonomous systems, accurate recognition of humans in real time and tracking of their trajectories becomes an achievable task. Nevertheless, algorithms are expected to perform well in a variety of challenging situations, such as in different weather conditions, light-cycle changes, occluded areas and crowded places. This adds complexity to the task, making it an on-going research topic. Task: Detect pedestrians in videos from outdoor surveillance camera and track trajectories of the moving objects.


A pre-trained detector implemented in Matlab Computer Vision Toolbox was used. The detector uses Aggregated Channel Features algorithm, in which both edge-based and color-based feature descriptors are combined. The aggregated channels used for classification are normalized gradient magnitude, histogram of oriented gradients and LUV color channels. The detector was trained with two different models (INRIA and Caltech).


The tracking algorithm is composed of three main steps: feature extraction, feature matching and trajectory update. For feature extraction, we use colour information from HSV (Hue, Saturation and Value) colormap. Each pedestrian object is split in three stripes and for each stripe histograms of every channel are calculated. Once we obtain feature vectors of pedestrians in current frame, we compute the affinity matrix between these vectors and previous ones. A cost function that represents similarity between features is computed. It is composed of the correlation coefficients between the features and a pedestrian position-dependent term, meaning that those potential matches that are spatially further away are more penalized.


4 Members from 4 different cultures and Education background!


Mountain View 7


Mountain View

Anđela Šaletić

Mountain View


Mountain View