An Experimental Study on Vision-based Multiple Target Tracking

Multiple target tracking has been an interested topic of research for vision-based traffic monitoring application because of its importance in associating multiple detected vehicles from consecutive frames of video. Before tracking multiple vehicles across frames, target detection algorithm, such as background subtraction is responsible for capturing the position of moving target in every frame. Tracking algorithm uses the measurements from the detection stage to relate the moving targets from previous frame with the current frame. Due to the limitation of performance in the target detection algorithm, it is not reliable to solely depend on the measurements computed from the detection stage. Thus, Kalman filter model has been adopted to compensate the fluctuation and missing measurements whenever the detection stage fails. The missing measurements are predicted based on the center position of vehicle and velocity estimation from the displacement of vehicle. Experimental study has been conducted on the vehicle tracking at road junction. The results showed that Kalman filter model assure the continuous tracking of multiple targets even though there are several lost measurements. Index Terms – Kalman filter, vehicle tracking.