Tracking interacting targets in multi-modal sensors
Object tracking is one of the fundamental tasks in various applications such as surveillance, sports, video conferencing and activity recognition. Factors such as occlusions, illumination changes and limited field of observance of the sensor make tracking a challenging task. To overcome these challenges the focus of this thesis is on using multiple modalities such as audio and video for multi-target, multi-modal tracking. Particularly, this thesis presents contributions to four related research topics, namely, pre-processing of input signals to reduce noise, multi-modal tracking, simultaneous detection and tracking, and interaction recognition. To improve the performance of detection algorithms, especially in the presence of noise, this thesis investigate filtering of the input data through spatio-temporal feature analysis as well as through frequency band analysis. The pre-processed data from multiple modalities is then fused within Particle filtering (PF). To further minimise the discrepancy between the real and the estimated positions, we propose a strategy that associates the hypotheses and the measurements with a real target, using a Weighted Probabilistic Data Association (WPDA). Since the filtering involved in the detection process reduces the available information and is inapplicable on low signal-to-noise ratio data, we investigate simultaneous detection and tracking approaches and propose a multi-target track-beforedetect Particle filtering (MT-TBD-PF). The proposed MT-TBD-PF algorithm bypasses the detection step and performs tracking in the raw signal. Finally, we apply the proposed multi-modal tracking to recognise interactions between targets in regions within, as well as outside the cameras’ fields of view. The efficiency of the proposed approaches are demonstrated on large uni-modal, multi-modal and multi-sensor scenarios from real world detections, tracking and event recognition datasets and through participation in evaluation campaigns.
- Theses