Person Counter

Count the total people present in the video (sequence of images). For a given frame, count of people in current frame, maintain the identify in further frames

Data set

MOT16 dataset obtained from website and paper

Initial setup

• MOT 16 dataset downloaded at $MOT16
• Install Tensorflow Object Detection API used for IoU
• Refer misc folder for auxillary files

Approach 1 : Position based (IoU)

For data association between two frames, use IoU of bb
Use groundtruth bb python person_counter.py --video 10 --iou 0.3 stepSize 1 --useGT
Use detection bb python person_counter.py --video 10 --iou 0.3 stepSize 1 --model 2
where

• video : video number of MOT16 dataset
• iou : IoU threshold for matching
• stepSize : frame processing rate (skips intermediate frames)
• useGT: use groundtruth boundary box (only for MOT16 train videos)
• model : model to be used for object detection

Approach 2 : Detect and track (OpenCV)

• A Detect-and-Track simulation framework where detection is done periodically and subsequent frames (window size) are tracked.
  Eg: Simulation on tracking window of size 3 frames [D T T T] [D T T T] [D T T T]
• Data association via IoU is performed on between windows
• OpenCV Trackers multiple object tracker

Usage

Conda environment file misc/cv.yml

1. Run simulation

python tracker_person_counter.py -v MOT16-10 -dh $MOT16 --tracker csrt 2>&1 | tee output/log/simulation.log

where

• dataset_home: path to dataset home
• video : video stream. e.g: MOT16-10
• tracker : Choose from csrt, kcf, boosting, mil, tld, medianflow, mosse
• window_size : window size (#frames) for tracking
• detect_speed: detection speed (sec)
• irs: infinite resource setting (do not skip any frames). Default resource constraint setting
  Note: Special cases
• Infinite resource setting: No frame are skipped. Use -irs flag which sets detection speed and tracking speed to 0
• No tracking: Perform only detection and data association via IoU. Use --window_size 0

2. Data association Generate global person id from detect-track paths

python pc_data_association.py -v MOT16-10 -dh $MOT16 -i output/local_track -o output/global_track

where

• dataset_home: path to dataset home
• video : video stream. e.g: MOT16-10
• input_dir : directory of detections having local id
• output_dir: detection to save detections after assigning global id

3. Evaluation
   Multiple object tracking metrics to evaluate the simulations.

python pc_mot_evaluate.py -dh $MOT16 2>&1 | tee output/log/mot_evaluate.log
./parse_log.sh output/log/mot_evaluate.log

• Setup mot metric pip install motmetrics
• Read about ID measures

Directory Structure

Dataset

MOT16
|-- train
   |-- MOT16-02
      |-- seqinfo.ini
      |-- img1
      |-- gt
         |-- gt.txt

Output

PersonCounter
|-- Output
   |-- local_track     // csv files having detect and track
   |-- global_track    // csv files after data association
   |-- log             // log files of the execution
   |-- data            // dataset structure with filtered GT, for pymotmetric evaluation
   summary.csv         // consolidated results from simulation

Convert frames to video

Convert sequence of images to video using ffmpeg
ffmpeg -framerate 7 -f image2 -i Frame_%03d.jpg ../output.mp4