thanks to Sirisian wrote, i adopted it a bit to my code for proximity sensor for neural network:

// convenient mapping for neural network distance sensor;
scalar get_distance_inverse_squared(const point& qp, const point& d, const segment& s)  {
    auto s0s1 = s[1] - s[0];
    auto s0qp = qp - s[0];
    auto dd = d[0] * s0s1[1] - d[1] * s0s1[0];
    if (dd != 0) {
        auto r = (s0qp[1] * s0s1[0] - s0qp[0] * s0s1[1]) / dd;
        auto s = (s0qp[1] * d[0] - s0qp[0] * d[1]) / dd;
        if (r >= 0 && s >= 0 && s <= 1) {
            // inverse square of distance, always less than 1.0
            return scalar(1) / (r*r + 1);
        }
    }
    return 0; // infinitely far, parallel, etc, no signal
}

thanks to Sirisian wrote, i adopted it a bit to my code for proximity sensor for neural network:

// convenient mapping for neural network distance sensor;
scalar get_distance_inverse_squared(const point& qp, const point& d, const segment& s)  {
    auto s0s1 = s[1] - s[0];
    auto s0qp = qp - s[0];
    auto dd = d[0] * s0s1[1] - d[1] * s0s1[0];
    if (dd != 0) {
        auto r = (s0qp[1] * s0s1[0] - s0qp[0] * s0s1[1]) / dd;
        auto s = (s0qp[1] * d[0] - s0qp[0] * d[1]) / dd;
        if (r >= 0 && s >= 0 && s <= 1) {
            // linear inverse of distance, always less than 1.0
            return scalar(1) / (r + 1);
        }
    }
    return 0; // infinitely far, parallel, etc, no signal
}