Week 12:
Machine learning concepts really helpful to improve the security and find the threats with least manual work. Deep learning and zero shot learning are so far have very low rates of false alarms and these should be applied. It can be seen that only one approach is not good but if we combine at least two techniques that it provides more better results like ANN and DBSCAN.
Week 11:
The above image illustrating the difference between normal and abnormal behavior. The main concept of problem domain of the project title is differentiate among normal and abnormal event. So, main artificial concepts are being applies and they are purely based on machine learning. But one more concept is social force model which finds the anomaly but calculating the distance between people. If the distance is dramatically increasing or decreasing, there can be anomaly.
Week 10:
It works in three levels which are bottom, intermediate and semantic. It works with two algorithms first one is ANN that one detects group of people whether DBSCAN works for individuals.
Week 9:
The image above shows the zero shot learning approach for anomaly detection. This approach does not require attributes. All it just needs detailed description. It will guess to find out the actual result.
Week 8:
The pic shown above is illustrating the concept of deep learning. input video is being segmented in frames for the further processing. Then features are extracted and object detection happens. After the classification it is decided that was there any misbehavior or not.
Week 7:
Entropy is another solution for the detection of anomaly. The above diagram shows the whole concept of Boltzmann entropy. if the heat energy in person jumps suddenly then it shows that object is moving. A sudden move could be anomaly or it shows person is running. If there is no change in energy it means its walking or idle.
Week 6:
Hybrid tracking is also a technique for crowd analysis and detection. This process works in three steps. first of all object is being tracked. Then the parameters of tracked person are obtained which are velocity, variance, direction of moving and entropy. Then eventually all these parameters are utilized to know the status of activity.
Week 5:
The picture above shows the concepts of deep multiple instance learning. Here 2ooo videos of different but both normal and anomalous events. After segmentation of both types of videos, a trained neural network will extract the features and then computes the ranking loss for both positive and negative bags.
Week 4:
Above image shows three types of crowd. First one is large size crowd, middle one is medium and right image shows small size crowd. Large size crowd is bit difficult to find anomaly because there are too many patterns are being followed.
Week 3:
Week 3 i spent to make the project proposal. I found some frameworks of crowd surveillance and behavior detection process which are following:
Week 2:
| NAME: | Muhammad Khurram Latif | ||
| PROJECT TITLE: | Crowd Surveillance and Abnormal Behavior Detection System | ||
| WEEK NO: | 2 | DATE: | 30-07-2020 |
| PLANNING | |||
| MILESTONE: | PLANNED: | ACTUAL: | COMMENT: |
| Find relevant data for research project. | 10 research paper or journal articles. | ||
| ISSUES: | |||
| DESCRIPTION | DATE: | ACTION/RESULTS | FINISHED (Y/N) |
| Topic selection from a very wide range of areas in IT. | 30-07-2020 | Collected some related data | Y |