Bike detection on Google Coral TPU

Real-time bike detection, tracking and counting system designed for Raspberry Pi with Google Coral TPU.

The system allows user to mark a virtual line segment on video input, that is used to count the number of bicycles that pass through it and identify their direction of crossing.

Accuracy Tests

Tests were performed for the following pre-trained models:

• Yolov5m - running on a GPU "best" (reference),
• Yolov5s - running on a GPU "smallest",
• SSDLite MobileDet - running on Google Coral TPU,
• EfficientDet-Lite1 - running on Google Coral TPU.

Detection Accuracy Calculation

When determining the level of detection, two sources of errors are considered:

• $\varepsilon_m$ - missing a vehicle by the system (number of vehicles missed),
• $\varepsilon_f$ - detecting a non-existent vehicle by the system (number of falsely detected vehicles).

If $N$ is the real number of vehicles that passed through the measurement point, the detection level is defined by the formula:

$$ r_d = \frac{(N - \varepsilon_m - \varepsilon_f)}{N} $$

Confidence Interval

The actual value of the tested parameter $p$ is higher than the value $\hat p_L$ with $95%$ probability.

$$ \hat p_L = \max\left\lbrace 0, \frac{2N\hat p + z^2 - \left[z\sqrt{z^2 -(1/N)+4N\hat p (1-\hat p) + (4\hat p - 2)} + 1 \right ]}{2\cdot(N+z^2)}\right\rbrace $$

• $\hat p_L$ - lower value of the symmetric confidence interval, calculated using the Wilson method,
• $\hat p$ - estimation of the given tested parameter,
• $z \approx 1.6448536$ - value resulting from the adopted confidence level (in this case 95%).

Results of the Accuracy Tests

Test data:

1. 4 locations,
2. 5 sessions (result calculated from the sum of the number of bicycles and errors for all sessions),
3. 362 bicycle passes.

Results:

	yolov5m	yolov5s	efficientdet_lite1	ssdlite_mobiledet
$$\hat p_L$$	0.8459	0.7923	0.6878	0.5328

Usage

python3 bikedet.py \
    --input <input_file_path> \
    --output-dir <output_dir_path> \
    --show-vid \
    --save-vid \
    --crop

Optional args:

    --model <model_path> \
    --labels <labels_path> \
    --config <config_path>