Model Based Technique for Vehicle Tracking in Traffic Video Using Spatial Local Features

In this paper, we proposed a novel method for visible vehicle tracking in traffic video sequence using model based strategy combined with spatial local features. Our tracking algorithm consists of two components: vehicle detection and vehicle tracking. In the detection step, we subtract the background and obtained candidate foreground objects represented as foreground mask. After obtaining foreground mask of candidate objects, vehicles are detected using Co-HOG descriptor. In the tracking step, vehicle model is constructed based on shape and texture features extracted from vehicle regions using Co-HOG and CSLBP method. After constructing the vehicle model, for the current frame, vehicle features are extracted from each vehicle region and then vehicle model is updated. Finally, vehicles are tracked based on the similarity measure between current frame vehicles and vehicle models. The proposed algorithm is evaluated based on precision, recall and VTA metrics obtained on GRAM-RTM dataset and i-Lids dataset. The experimental results demonstrate that our method achieves good accuracy.