An accuracy boosting tool of video analytics by object tracking with motion vector in the bitstream.
python 3.7
ffmpeg version 4.4.1 Copyright (c) 2000-2021 the FFmpeg developers
built with gcc 7 (Ubuntu 7.5.0-3ubuntu1~18.04)
configuration: --enable-shared --disable-static
install imutils and cv2
Before running the code, make sure ffmpeg-4.4.1 is installed, if not, do:
1.By https://git.ffmpeg.org/ffmpeg.git, get ffmpeg-4.4.1 version (or 4.4.x)
2.Do this in turn to install ffmpeg
cd ffmpeg-4.4.1
./configure --enable-shared --disable-static
make -j8
make install
Now, with the repo folder h264_cabac. C will libavcodec folder of h264_cabac. C :
cd ffmpeg-4.4.1
Compile：
make
make install
Execute：
ffmpeg -c:v h264 -i input.mp4 -f null -
Subsequently, in the execution of the instruction folder to obtain the mv.txt file to save the motion vector information of all frames;
Detect input.mp4 using any object detection model, the test results are saved in the ```results`` CSV file.
Get mv. TXT file, and get to the test results results CSV file, will redefine the inspection results for dds_utils. Py 在 Region structure, Run the process_results.py file to reuse the results and save the final results as a final_results CSV file.
class Region:
def __init__(self, fid, x, y, w, h, conf, label, resolution,
origin="generic"):
self.fid = int(fid)
self.x = float(x)
self.y = float(y)
self.w = float(w)
self.h = float(h)
self.conf = float(conf)
self.label = label
self.resolution = float(resolution)
self.origin = origin
The original result results and the reused final_results as nomv_filename and 和 filename variables of visual.py file, after execution, a visual video file can be obtained. For example, demo.mp4 in the repo.
In the following picture, the green box is the detection result of one frame per second by business process, and the blue box is the MV-based tracking result:
