Aiming at the problem of sparse and low information content of LiDAR data,which makes it difficult to recognize human features,a real-time detection method for human legs based on convolution attention mechanism is proposed.The laser point information is preprocessed by a depth-guided sliding window,so that the object has the same feature information at different distances.Through aggregation of time information,more abundant spatial representation of LiDAR data can be obtained,which reduces the operation time.The characteristics of the associated dislocation with spatial neighborhoods are analyzed by the convolution attention module and the autoregression model.To verify the detection effect of the proposed algorithm on pedestrian legs,under three evaluation radius of DROW validation set,area under the curve(AUC)is increased by more than 21%,F1 is increased by more than 14%,and detection time is reduced by 13 ms on average.The experimental results show that this algorithm has higher detection precision and faster operation speed than DROW algorithm.
关键词
一维卷积神经网络/注意力机制/二维激光雷达/人体腿部识别
Key words
1D-CNN/attention mechanism(AM)/2D-LiDAR/human leg detection
维普
万方数据