You are on page 1of 3

113

CHAPTER 7

CONCLUSION AND FUTURE SCOPE

Significant progress has been made in video object path prediction


and tracking during the last decade. Numerous efficient trackers have been
developed and deployed which can track objects in simple real- time. However,
the inference from the literature shows that the conventions used in those works
make the prediction and tracking a simpler one. In many realistic scenarios,
motion smoothness, object occlusion, illumination variation are exploited and
limits the usefulness of the object tracker. Thus prediction and tracking of
dynamic objects and its associated problems of video segmentation, object
detection are very active areas of research and researchers continuously put forth
enormous solutions for real-time object tracking

In this dissertation, foreground object prediction and tracking is


achieved by developing efficient algorithms for video segmentation, object
detection, object path prediction and tracking. Basics of video segmentation are
introduced in chapter 3 with some improved metrics to reduce the processing
time for analyzing a video. Instead of focusing on the frame-to-frame prediction
of moving objects, a simple color histogram based threshold technique is
proposed to identify the candidate frames and hence the number of frames to be
processed gets reduced.

Semantic information about the objects present in an image is


basically required for meaningful segmentation. Keeping this in mind, the
proposed framework employs superpixels for segmenting the foreground
objects. Instead of processing the whole image, small clusters of pixels get
114

processed. To segment the moving object from frames with good boundary
adherence, a simple segmentation algorithm SUBBG is proposed in chapter 4.
Motion pixels are identified and grouped as clusters to segment the foreground
objects from the frames. A segmentation framework with SOCRA algorithm is
employed to optimize the CRFs for efficient foreground segmentation.

In chapter 5, conventional object tracker Kalman filter is amended to


optimize the trail path of the object with improved metrics in the epipolar
geometry. The objects present in consecutive candidate frames are far away than
the objects present in consecutive video frames. Hence the search region for
predicting the object centroid is expanded by varying the rotation angle. The
location of foreground objects in various orientation bins is predicted
successfully by the amended Kalman filter.

In the final part of the dissertation, paths of the dynamic objects are
generated and tracked using the proposed H-track algorithm. In general,
Hungarian algorithm is employed to locate and assign the object centroids to the
predicted tracks. The cost matrix of the Hungarian algorithm is improved with
the IoU and Euclid distance measure between the detected objects and predicted
objects.

The proposed system can be extended to build intelligent monitoring


systems for various surveillance applications. Activity pattern of the detected
objects can easily be inferred from the detected tracks. The learned patterns are
then used by the surveillance systems to analyze the future activity of the
objects. Since continuous frame sequences are not employed, appearance and
disappearance of the same object due to occlusion badly affect the success rate.
Main drawback of the proposed algorithm is that the videos with heavily
crowded moving objects may result with multiple overlapping of object clusters
during segmentation. To overcome the above drawback additional features
115

(Shape, texture) can also be included in the inference network in future. In the
future, semantic labeling of moving objects will be done to overcome the false
alarm and miss detections.

You might also like