Design and Implementation of Leg Detection System Based on Laser Scan Data for Robot Perception

This paper designs and implements a leg detection system based on laser scan data. People detection is an important function for autonomous navigation robots, especially in strongly interactive application scenarios. Though it's more intuitive to detect people by means of computer vision, it often means more computation, higher latency, and higher hardware costs. In this work, we propose an approach to detect human leg using pure two dimensional laser scan data with higher accuracy. Our system is based on a random forest model for human leg detection. Different from the original solutions, we add a series of feature filtering mechanisms to filter the leg candidates based on their data characteristics. The features will be put into the random forest model after filtering, and the confidence of being human legs will be calculated. Then, we verify the coordinates of the human legs using the map masking mechanism. It verifies the coordinates of the human legs with map data, which greatly reduces the probability that cylindrical objects such as pillars in the environment being misidentified as human legs, removing most of the false positives provided by the random forest model.