1. TECHNICAL FIELD OF THE INVENTION

The present invention relates generally to an object tracking system and method which is able to track multiple objects by locating the objects in the current image sequence and assigning a consistent object identifier to each of the objects throughout the image sequences, said system and method integrates the approach of bottom-up and top-down.

2. BACKGROUND OF THE INVENTION

Object detection and tracking is important in many applications such as in visual surveillance, video communication and human computer interaction. Object detection and tracking can be defined as the process of locating a moving object of interest in the scene and assigning a consistent object label to the same object throughout the image sequences. This process is not a trivial task due to several difficulties. The said difficulties are fast moving object, scene illumination changes, image appearance changes as the image viewpoint changes, partial or full occlusion of object- to-object and object-to-scene, non-rigid object shape as it moves and real time processing requirement. Methods of visual tracking system can be divided into two major approaches: bottom-up and top-down. In a bottom-up approach, the image is generally segmented into objects which are then used for tracking. Some common bottom-up algorithms are blob tracking, contour tracking and visual feature matching. Said bottom-up approach has mostly low computational complexity but not robust enough to track partial or fully occluded object. On the other hand, the latter approach is mostly associated with high computational complexity but more robust to track occluded object. This top-down approach basically generates object hypothesis and tries to verify them using image. The object to be tracked is sometimes initialized manually by the user or determined automatically using some intelligent and complex object detection method. The said initialization stage is critical for this approach and it will affect the performance of the tracking process. Some common tracking algorithms for said top-down approach are Kalman filter and Particle filter.

It would hence be extremely advantageous if the above shortcoming is alleviated by having an object tracking system and method which is able to track multiple objects by locating the objects in the current image sequence and assigning a consistent object identifier to each of the objects throughout the image sequences, said system and method integrates the approach of bottom-up and top-down. 3. SUMMARY OF THE INVENTION

Accordingly, it is the primary aim of the present invention to provide an object tracking system and method which combines the advantages of both the top-down and bottom-up tracking approach to complement the limitations of both said approaches.

It is yet another object of the present invention to provide an object tracking system and method which has low computational complexity.

It is yet another object of the present invention to provide an object tracking system and method which is able to track partially or fully occluded objects.

Other and further objects of the invention will become apparent with an understanding of the following detailed description of the invention or upon employment of the invention in practice.

In a preferred embodiment the method of tracking objects using a tracking system comprising steps of: i. sensor and camera performing motion detection on the ^current image frame to obtain current motion map in binary format indicating location of motion and non-motion pixels; ii. any acceptable processing means performing pre-processing on said current motion map by eliminating noises and labelling said motion map; characterised in that iii. any acceptable processing means performing overlapping region analysis on current and previous motion map to determine the relationship between said current and previous motion pixel groups in the said motion map; iv. any acceptable processing means invoking one of a plurality of blob events based on the result of said overlapping region analysis.

According to a another embodiment of the present invention there is provided, An object tracking system for detecting and tracking multiple objects by localizing said objects and assigning consistent label to each object throughout a sequence of images, comprising:

i. any acceptable sensoring means;

ii. any acceptable processing means;

iii. any acceptable image capturing device; characterized in that

said any acceptable processing means are configured to process image from said image capturing device characterized in that said processing means is configured to perform a method according to said preferred embodiment.

4. BRIEF DESCRIPTION OF THE DRAWINGS

Other aspect of the present invention and their advantages will be discerned after studying the Detailed Description in conjunction with the accompanying drawings in which: FIG 1 shows a flow chart of an object tracking method which comprises an integration of the bottom-up and top-down tracking approaches.

FIG 2 shows the rule of determining events based on the overlapping previous-current blobs test.

FIG 3 shows a flow chart of the sub-steps of the step of Perform Process New Object in FIG l.

FIG 4 shows a flow chart of the sub-steps of the step of Perform Top- Down Tracking Approach in FIG 1. FIG 5 shows a flow chart of the sub-steps of the sub-step of Alignment Process to Align Estimated Object Location and the Motion Blob in FIG 4.

5. DETAILED DESCRIPTION OF THE DRAWINGS

In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the invention.

However, it will be understood by those or ordinary skill in the art that the invention may be practised without these specific details. In other instances, well known methods, procedures and/ or components have not been described in detail so as not to obscure the invention. The invention will be more clearly understood from the following description of the embodiments thereof, given by way of example only with reference to the accompanying drawings which are not drawn to scale.

Referring to FIG 1, there is shown a flow chart of an object tracking method which comprises an integration of the bottom-up and top-down tracking approaches. In the bottom-up approach, a blob based tracking method is proposed as the main tracking sub-system. The idea is to track each of motion blobs appeared in the current frame by using at least one tracker. Each tracker stores information of object features such as colour and texture, object centroid and area and object identifier. Firstly, a motion map of current frame is obtained by applying motion detection process (102) to the current image frame (104). Motion map is a binary map which indicates the location of motion and non-motion pixels. A group of motion pixels are called motion blobs. Said motion map will be pre- processed (106) to eliminate noise, fill up holes and label each of the motion blobs. Next, said processed motion map from the current frame will be compared (108) with its equivalent motion map from the previous frame (110) to obtain previous-current motion blobs relationship. Each motion blob in the current motion map will be compared to other motion blobs from the previous motion map to determine one of a plurality of events comprising: new object (112), existing object (114), blob splitting (118) or blobs merging (116). In this case, overlapping region analysis (108) is used as a comparison method.

Referring now to FIG 2, there is shown the rule of determining events based on the overlapping previous-current blobs test (108). Blob events comprises of any events of new object (112), objects merging (116), objects splitting (118) and existing object (114). If one motion blob in the current motion map overlaps with only one motion blob from the previous motion map (122), the current motion blob will be given the same object label as the previous motion blob. Given another example, if one motion blob in the current motion map overlaps with more than one motion blob in the previous motion map (124), it is interpreted that more than one motion blob in the previous motion map has merged with the current motion map, thus the event of object merging (116) is detected. In another case, if more than one motion blob in the current motion map overlaps with only one motion blob in the previous motion map (126), it is interpreted that the previous motion blob is being split in the current motion map, thus the event of objects splitting (118) is detected. Using another example, if any motion blob in the current motion map does not overlap with any motion blob in the previous motion map (120), the motion blob in the current motion map is interpreted to be a new object, thus the event of new object (112) is detected.

Each of the detected events will invoke different processes of updating the tracker information and assigning the object identifier. The processes are described as follows: i. if one previous blob is related to only one current blob (122), then the current blob is corresponding to an existing object being tracked and it will be continually tracked and labelled as the same object label assigned to its related previous blob; ii. if the current blob is not related to any previous blob (120), that blob corresponds to a new object to be tracked and the sub-step of Process New Object (112) will be invoked; iii. if one previous blob relates to more than one current blob (126), the sub-step of Splitting Event (118) is detected; iv. if more than one previous blob relates to only one current blob (124), the sub-step of Merging Event (116) takes place and top- bottom tracking approach will be invoked.

Referring now to FIG 3, there is shown a flow chart of the sub-steps of the step of Perform Process New Object (112) in FIG 1. Initially, ratio test is performed on the motion blob (302) identified as new object to be tracked. The action of ratio test is done on the motion blob bounding box enclosing the said motion blob. Ratio test comprises of the calculation of the ratio of the height by width of the said motion blob bounding box. This ratio test is to determine whether the blob contains single or multiple objects (304). If the blob contains multiple objects, human detection algorithm is applied on the blob (306) to separate multiple objects into each single object. Only a complete single object (track-able object) within the multiple objects blob will be tracked at this stage. Then, for each single object blob, its object properties will be extracted (308) together with its location information (centroid and enclosing bounding box coordinates). Colours and textures properties are also used to represent the object as they are robust to partial object occlusion, rotation, scale invariant and computationally inexpensive. After that a tracker is initialized for each of single track-able objects within the blob (310). Each tracker will hold the information of object features and location properties. In order to assign object label identifier, the object features first need to be matched with all objects in the memory of the said tracking system (312). Objects in said memory are the objects have been tracked before and already leave the scene for not more than a certain number of frames. If the current object is found similar to any object in the memory (314), then current object is assigned as the same identifier as the object in the memory (316) and the object in said memory will be removed (318). If the current object is found to be different from any object in the memory, a new object identifier will be assigned to current object (320).

Referring now to FIG 4, there is shown a flow chart of the sub-steps of the step of Perform Top-Down Tracking Approach (116) in FIG 1. In this top-down approach, particle filter based object tracking is employed. Initially, said particle filter tracker corresponding to each of the objects in the merged blob will be initialized (402). The step of initializing particle filter object tracking method comprises of said acceptable processing means performing the following sub-steps: i. extracting label and location for object from the previous image frame; ii. extracting object properties from said object location identified in the previous image frame; iii. storing said object label, location and properties in an object tracker data structure.

In the initialization process, target model features are generated using the object colour and texture information from the previous frame in which the object appeared as a single object just before it merges in current frame. During the localization process (404), the said tracker will find the most probable sample distribution in the current frame by comparing the target model with the current hypotheses of the particle filter using Bhattacharyya coefficient, which is a popular similarity measurement method between two distributions. After all the objects within the merged blob are being localized by their own trackers, their predicted locations will then have to be aligned with the motion blob area (406). Referring now to FIG 5, there is shown a flow chart of the sub-steps of the sub-step of Alignment Process to Align Estimated Object Location and the Motion Blob (406) in FIG 4. The purpose of the alignment process is to fine-tune the object location in the current frame as predicted by said particle filter by incorporating the information from the motion blob. In this process, the percentage of overlapping area between each of predicted object location as first outlined by an enclosing bounding box and its motion blob area is evaluated (502). This value of overlapping percentage determines which object location needs to be readjusted. If the overlapping percentage is less than a certain threshold value (504), then the estimated object location will be readjusted. Subsequently for each object which requires adjustment, the possible moving directions so as to move the predicted object bounding box to be within the motion blob area at least more than a certain threshold value is determined (506). These possible directions are restricted to be among the 8-neighbor directions.

Assuming the imaginary line connecting the predicted object centroid to the point where the predicted object centroid has to move into the motion blob denotes as a moving line. Along the possible moving line, the object features are extracted (508) and similarity measurement between the said extracted features and the object model features is calculated (510) and the adjusted location of the object corresponds to the location along the moving line with highest similarity value (512). While the preferred embodiment of the present invention and its advantages has been disclosed in the above Detailed Description, the invention is not limited thereto but only by the spirit and scope of the appended claim.