On the Essence of Unsupervised Detection of Anomalous Motion in Surveillance Videos

Below are the dense trajectories extracted for anomaly detection datasets. These trajectories were extracted using IMPROVED TRAJECTORIES implementation of Wang et al. paper "Action Recognition with Improved Trajectories".

[UCSDped1]         [UCSDped2]        [UMN]          [PETS2009]

Once you unzip the .rar file, you will get an array of trajectories (a binary file in NumPy .npy format) for each test video. Each array of trajectories has a size of N x L x M, where N is the number of trajectories, L is the trajectory length (set to 15), and M is the frame number along with x,y coordinates.

To read the .npy files in Python, you can use the following line of code that reads the file and saves it in NumPy array:

			import numpy as np
			test001_arr=np.load('Test001.npy')
			

To download the dataset test videos click on the following dataset name and it will direct you to the website:
[UCSD. Anomaly Detection Dataset]
[UMN. Detection of Unusual Crowd Activity]
[PETS2009. Flow Analysis and Event Recognition]

REMEMBER:
In UCSDped1 videos, humans and other objects move toward or away from the camera resulting in noticeable perspective distortion. As a result, objects nearer to the camera appears to move faster than those further from the camera ven though they may move at the same actual speed. To overcome this distortion, the feature points are projected onto the ground plane using an estimated homography. Then, the speeds of the feature points are computed after projection. The following are: the estimated Homography (3x3 matrix) for UCSDped1 videos, an example Python code that projects the points into ground plane, and a test video used to run the code.

[UCSDped1_H.npy]            [projectIntoGP.py]            [Test001.avi]