Object Tracking Clause Samples

Object Tracking. The output from the robust object detection is passed to a tracking module in order to perform robust and long term tracking, [cf. Objective 1]. Here the primary aim was to maintain robust tracks for all the individuals in the scene throughout the sequence in order to provide input into the higher level analysis such as threat analysis. Robust tracking refers to the real-time generation of reliable “tracklets”, while the long- term tracking refers to the processing of these tracklets to form complete tracks. The tracking module had two main roles: to produce a spatio-temporal detection of relevant entities – a track – and assign a unique ID tag to each track. As the track detection and ID track maintenance functionalities are closely intertwined, these aspects were jointly designed. 1. The final development tasks looked at the design of functions for long term track association. As the primary output from SUBITO was a customer demonstration, it was felt important that the system be able to run as near real-time as possible. Current state-of-the-art tracking algorithms, involving sophisticated detection steps, hardly reach a computational rate of a few HZ. The approach adopted for SUBITO was to initially define a baseline running processing chain that would operate real-time and to which performance improvements could be made incrementally, as follows: • The baseline approach was a simple scheme using Kalman filter models and ▇▇▇▇▇▇▇ associations to provide tracking for detections obtained using simple background subtraction techniques. • This baseline was then modified to provide enhanced localisation in space for tracked objects by exploiting multi-view detections to improve the eventual threat detection and reduce the false alarm rate. • The implemented multi-view algorithm used homographic constraints and a priori hypotheses to produce an estimate of the 3D bounding box containing the moving individuals. • Extra spatial reasoning, using accumulation maps, was implemented to filter out some of the artefacts generated in fusing the data from the multiple views, which proved especially fruitful when the scene contained multiple people interacting, Figure 2-9. More details can be found in Ref. 13 • Further performance improvement was produced by introducing better track and ID management by maintaining several concurrent track hypotheses in a delayed decision making framework based on the Multiple Hypothesis Tracker (MHT). This scheme better manages small o...